How Gibraltar's systems work

What it is

Everyone signs in to the ops dashboard with their work email and a password of at least 12 characters. New people get an emailed invite link and choose their own password. Anyone already signed in to the Sales CRM is recognized by the dashboard automatically (single sign-on, meaning one sign-in covers both).

Why we built it

The dashboard holds the FD lending ledger, the P&L and buyers' personal details, so it is built to 'fail closed': a person with no tabs sees nothing. The invite flow means no admin ever knows or texts anyone's password.

How it works

An admin creates the login in Users & Access. The system then emails a one-time link from anthony@ that works once and expires after 7 days. The person opens it and sets their own password. An admin can instead set a temporary password, which expires after 7 days. The first sign-in then forces a new password and shows a 'save your password' screen and instructions for adding the app to a phone. A sign-in lasts 30 days. In a normal browser tab it ends when the tab closes; in the installed app it is remembered. Single sign-on: the front door reads the CRM sign-in cookie and passes the person's email to the server, which matches it to a dashboard user. The CRM sign-in only proves who you are. Your tabs still come from the dashboard's own user list.

How you use it

Go to ops.gibraltarbuilders.info and sign in. There is no self-service 'forgot password'. Ask an admin (Anthony, or Daniel as an admin) to send a new invite from Users & Access, which also works as a password reset.

Who has access

9 dashboard logins: the shared 'admin' login (Anthony's), Daniel, Rich, Tim, Alessandra, Kelsey, Kody, Jacob and Jim. Kody and Jacob have never signed in (their invite links expired Oct 5). Jim has not signed in yet; his temporary password expired this morning, and a new invite went to him today and is good until Oct 23.

Watch out: Anthony has no named login of his own, so his changes are recorded as 'admin'. Daniel has no password sign-in on record in the last 30 days and most likely comes in through his CRM sign-in. Repo notes from August say the key shared by the CRM and the dashboard for single sign-on is overdue to be replaced with a new one. I could not verify this without reading the secret, so I did not.

What it is

Gibraltar's internal web dashboard. It is one website with a sidebar of tabs (Realtor Accountability, Off-Market Lots, Metrics, Progress, Homes & Stages, Stage Updates, Project Timeline, Trade Directory, AP Comms, Liquidation, E-signatures, Users & Access, Settings), a Finance group (FD Funding, P&L 2026), a Sales CRM group, and an Assistant.

Why we built it

To put the company's working tools and live numbers in one place with one login, instead of scattered spreadsheets and separate apps.

How it works

A Cloudflare front door, which sits in the 'thoughtsinform' Cloudflare account Daniel owns, serves the pages. It passes data requests to the ops server, a cloud server shared with Anthony's Hively platform, which keeps everything in one database. A change to the web pages goes live automatically when Anthony pushes it to GitHub. The system checks itself after each such update and marks the update failed if the site is not answering. Server changes are copied up by hand through the Mac Mini. Jobs that cannot run on the cloud server run on the Mac Studio: photo processing, the stage-update email sender, the daily lot and permit research, and the nightly backup. Every night at 2:30 AM the Studio backs up the ops database and the CRM database. Today's backup succeeded.

How you use it

Open ops.gibraltarbuilders.info in any browser, or use the installed app. Which tabs you see depends on your grants.

Who has access

Everyone with a dashboard login; tabs vary per person (see each entry).

Watch out: The ops server is still on Anthony's Hively server under a Hively web address. Moving it to Gibraltar's own server is an open 'later' item. An old, frozen copy of the dashboard's front end is still publicly served at an old address (left there on purpose on Sep 15). It still connects to the live server, though data still needs a sign-in. The old developer-digital.cc address no longer loads; ignore it.

What it is

Tracks the 697 buyer leads given to the sales agents: 453 bought and skip-traced, meaning looked up to find their contact details, plus 244 from Meta (Facebook and Instagram) ad forms. It checks county property records to see whether each lead later bought a home, and whether that purchase went through Gibraltar's agents or an outside agent. It has Overview, Buyers, Leads, Pipeline and Tasks views, and a Property Lookup sub-tab for checking any address.

Why we built it

To hold the sales team accountable for leads Gibraltar paid for, and to spot leads who bought elsewhere. Leads who have not bought are treated as buyers still in the market.

How it works

Every Friday a Mac Studio job checked county and paid property-data sources and pushed the results to the dashboard. The last push was Oct 9. That Friday job was switched off on Oct 9 because it could start paid data pulls and skip-traces with no per-run approval. Its paid sources were already out of credit or blocked, so it had been finding little new. The newest lead in this tab is from Aug 12. New leads now live in the Sales CRM, which has 1,021 contacts.

How you use it

Dashboard > Realtor Accountability (the home page for those who hold it).

Who has access

Admin-role holders only: Daniel, Rich, Tim and the shared admin login.

Watch out: Current figures: 1 sold through Gibraltar's agents, 3 through outside agents, 116 active and not bought, 379 pending, and 134 that could not be matched to a person. These will not change until the weekly check is switched back on. Anthony's 10/9 rule is that the paid version stays off until a spend guard exists. A free replacement is being built from Daniel's seat: on Oct 16 (5:50 to 5:59 PM) his Claude committed a check that uses only free county data to see whether a lead bought a Brevard home. Tested against the 17 real 2026 buyers in the Closings Log, it found 12; the misses were mostly recent closings the county had not posted yet. It is a dry run only: the weekly schedule still points at the paused paid job, a real run needs the Studio, and it cannot name the selling agent.

What it is

A database of about 8,500 vacant lots in Brevard, sorted into Ready to buy, Customs, Check first and Skip. It opens on a call list of about 1,200 lot owners, with a log for calls, offers, do-not-call requests and site checks.

Why we built it

Anthony's target is about 9 lots bought a month, roughly 100 in a year. The old score ranked unbuildable 'paper' lots highest, so the revamp ranks by real buildability instead: sewer or septic, flood zone, scrub-jay habitat (a protected bird whose habitat restricts building) and wetlands.

How it works

Every day at 6:30 AM a Mac Studio job refreshes owners, hazards and tax-deed status, then pushes lots and building permits to the server. Today's push ran at about 7 AM. The list puts city-water lots first and septic lots after them, flagged with a warning. An offer is refused until flood, wetland and scrub-jay checks are clear. Nothing in this tab dials, texts or emails anyone. Tax-deed data only comes from realTDM pages (the county's online tax-deed sale listings) that a person saves into a folder on the Studio, because realTDM blocks automated access.

How you use it

Dashboard > Off-Market Lots > Call list. Log each call or offer on the owner's row.

Who has access

Admin-role holders (Daniel, Rich, Tim, admin). No callers have logins yet.

Watch out: No calls, offers or site checks have been logged yet. The callers have not been named. The full hazard pass and the 'risk gates' (the switches that would enforce the hazard checks) are Anthony's to run. The paid Friday sweep stays off.

What it is

The lending ledger for FD Funding, the company that funds Gibraltar's lots and builds. 184 properties: 23 under construction, 32 not started, the rest closed. 478 advances totaling about $9.1M. Sub-tabs: Overview (with a one-page print), Ledger, Pending (closing documents found in Kelsey's mailbox) and Feeds & QuickBooks.

Why we built it

To replace Finance's manual tracking with a live ledger that computes fees and interest, and shows what is owed on each property.

How it works

Two-way sync with Finance's FD Funding Google Sheet: the server reads the sheet every 10 minutes, and edits made on the dashboard are written back to it. An email agent files closing documents from kelsey@ into Pending (the newest arrived today). Anyone holding this tab gets a desktop browser alert when a new closing document arrives. Interest and fees are calculated live, never stored. Following Alessandra's review, the ledger shows both the original lot a check was written for and the property where the money counts.

How you use it

Dashboard > Finance > FD Funding.

Who has access

Full access: Alessandra plus admin-role holders (Daniel, Rich, Tim, admin). View-only: Kelsey and Jim.

Watch out: The QuickBooks Desktop link shown under Feeds & QuickBooks has been dead since Gibraltar moved to QuickBooks Online in mid-September; the QuickBooks Online connection is not built. On Oct 16 Anthony marked the FD Funding ledger project 100% complete; the open item about two duplicate addresses on the FD sheet was closed as dropped from scope, not fixed. Alessandra and Kelsey used the dashboard today.

What it is

Gibraltar's 2026 profit and loss: Overview, Job Detail, Job Costing, P&L by Month, and a Checks panel that flags where the numbers do not tie out.

Why we built it

Finance asked for it, and the text that led to it said 'P&L 2026 is something separate from FD funding', so it is its own grant under a Finance heading. On Oct 14 Alessandra asked to link it directly to her workbook until QuickBooks Online is connected.

How it works

Every 15 minutes the server reads Alessandra's 'Gibraltar Builders_P&L_Q3' workbook from Google Drive and recalculates every total to the cent. It shows the numbers as she has them, plus the checks. If a read fails, the last good copy stays up. The latest read was Oct 15 and succeeded. Cross-checks against FD records appear only for readers who also hold FD Funding.

How you use it

Dashboard > Finance > P&L 2026. Alessandra keeps her workbook updated and the tab follows within 15 minutes.

Who has access

Full: Alessandra and admin-role holders. View-only: Kelsey and Jim.

Watch out: A named contractor's line is visible to anyone who can open the tab, which is still an open question for Anthony. The longer-term plan is to read QuickBooks Online directly (planned, not built).

What it is

A card for every home on the GB Inventory sheet: 67 active, of which 49 are at stage 0 and 7 at stage 10. Each card shows the build stage (1 to 10), stage dates, the stage checklist, photos, buyer and agent contacts, the superintendent, the buyer-portal link, and a Liquidation button.

Why we built it

To track every build in one place, feed buyer updates and the buyer portal, and trigger the liquidation program when a home nears completion.

How it works

Each home's stage follows the 'Current Phase' column of the GB Inventory sheet. Stage start dates come from the workbook's 2026 Production tab and sync every 10 minutes. An office user can also set a stage by hand. Buyers are filled in automatically from Kelsey's mail by the buyer harvester. Photos arrive from BuilderTrend once a day at 6 AM. Project managers and superintendents see only their own homes and can only add photos and tick the checklist.

How you use it

Dashboard > Homes & Stages. Open a card to change a stage, approve photos, create or email a portal link, or press Liquidate.

Who has access

Admin-role holders (Daniel, Rich, Tim, admin). Kody and Jacob have project-manager logins limited to their stage range but have never signed in.

Watch out: The stage checklists have no items at all. Jim still owes the 'blue-tape-ready' standard (the final punch-list check before closing) and Tim's stage-10 checklist. The sheet often lags reality, and the Homes card does not show the co-buyer yet.

What it is

The superintendents' Daily Log photos in BuilderTrend are copied into Homes & Stages and filed into each home's 'Stage Update Photos / Stage N' folder on the shared Google Drive.

Why we built it

BuilderTrend does not record which stage a photo belongs to. Anthony wanted the photos on each home, in the right stage folders, without anyone copying them by hand.

How it works

Every day at 6 AM a robot on the Mac Studio, signed in to BuilderTrend, reads the Daily Logs and pushes new photos to the server. The last run was today at 3:32 PM. Each photo gets a stage: an exact one from the sheet on the log date, or a suggested one from a free local image model, with a person approving or correcting it so the model learns. So far 1,361 photos have been synced, and 1,257 are filed in Drive. There is also a second, older path: photos dropped into Drive Stage folders are meant to flow into the dashboard. That path has failed every 15 minutes since Sept 18 because a Google sign-in on the Studio expired.

How you use it

Dashboard > Homes & Stages > open a home > Photos: approve, move to another stage, or mark 'Not a build photo'.

Who has access

Anyone with Homes & Stages

Watch out: None of the 1,361 BuilderTrend photos has been approved yet, and 104 are waiting for someone to pick a stage. The buyer portal only shows approved photos, so buyers see almost none. The Drive-folder intake needs Anthony to sign Google in again at the Studio, or the durable fix of moving it to a different Google login that does not expire.

What it is

A private web page per home that shows the buyer their build progress (a 10-stage ring), approved photos, timeline changes and their superintendent's contact details. Buyers never see a login screen.

Why we built it

To keep buyers informed without phone calls, and to look professional to clients.

How it works

The office presses 'Create link' on the Homes card. The link is shown once and only a scrambled copy is stored. 'Email the link' drafts an email that goes through Stage Updates for approval. The page always shows in dark colors on purpose, because it is client-facing.

How you use it

Homes & Stages > card > Create link (copy it), or Email the link, then approve the email in Stage Updates.

Who has access

Buyers holding a link. Office users issue links.

Watch out: 3 links are live, and the last time any buyer opened one was Sept 28. Only 4 buyers and 2 co-buyers are recorded on cards. Because outgoing buyer email is still in test mode (see Stage Updates), link emails do not reach buyers either.

What it is

Kody Arnold owns stages 1 to 4 and Jacob McCreary owns stages 5 to 10. The superintendent for a home's current stage is copied on every buyer email and shown on the Homes card and the buyer portal. Every Tuesday each superintendent is due an email listing their homes that need photos.

Why we built it

Anthony's Sept 28 ask: put the superintendents into the stages pipeline so buyers know who builds their home and photos keep coming.

How it works

A Superintendents panel inside Homes & Stages sets the stage ranges. The Tuesday 8 AM digest runs in test mode: 6 digests have gone out, all redirected to Anthony's test inbox, the last on Oct 13. It goes live only when both the panel switch and a server setting are turned on.

How you use it

Admins: Homes & Stages > Superintendents panel. Superintendents: sign in and see only their homes.

Who has access

Admins edit; Kody and Jacob see their own homes.

Watch out: Neither superintendent has activated his login, since the invites expired Oct 5. Turning the digest live is an open item waiting on Anthony to read a test digest first.

What it is

Reads Kelsey's 'UNDER CONTRACT' emails and the transaction coordinator's Transaction Summary PDFs, and fills in the buyer and co-buyer on the matching home card.

Why we built it

The sales side records buyers in email, not in the workbook. Anthony wanted the buyer known before Tim issues the portal link.

How it works

Every 15 minutes the server reads kelsey@'s mail (read-only) and matches each address to a home. It fills a buyer only where the card has none and no person ever removed one. It sends nothing. So far it has looked at 177 emails and filled buyers from 9; the last fill was Oct 5.

How you use it

Nothing: it runs by itself. Admins can see what it did inside Homes & Stages.

Who has access

Admins

Watch out: Four homes whose test buyers were removed by hand on Sept 28 (113 Crescent, 3014 Gabrish, 1387 Platt, 3145 Sharon) will never auto-fill until Anthony changes a rule. One home (2166 Santa Maria) has no buyer email in any source.

What it is

The approval desk for every email to a buyer: stage updates and portal links. You read the exact email, photos included, before you can press Approve.

Why we built it

Rule: nobody signs off on a client email they have not read in its final form.

How it works

A draft is created from a stage template on the Homes card. A person edits it and approves it. The Mac Studio checks for approved emails every 5 minutes and sends them from scheduling@ (the cloud server cannot send email directly). Right now the sender is in TEST mode, so every recipient is redirected to Anthony's test inbox. A newer sender on the server is installed but switched off.

How you use it

Dashboard > Stage Updates: read, edit, Approve, Discard or Retry.

Who has access

Admin-role holders (Daniel, Rich, Tim, admin). Approving buyer email is its own grant, separate from Homes.

Watch out: On hold since Sept 22 by Rich: 'stage updates on hold currently until we finalize it with Tim.' Only 3 updates have ever been made: 2 sent and 1 failed (1387 Platt Ave, which needs a person to re-draft it).

What it is

An internal planning board: projects as bars on a week-by-week timeline, with tasks, owners and progress. It includes an AI helper that drafts a plan you preview before saving.

Why we built it

A place for the team to plan its own work. Nothing in it is customer-facing.

How it works

Projects and tasks are stored on the ops server. Bars can be dragged. Progress can never show credit for days that have not happened yet.

How you use it

Dashboard > Project Timeline.

Who has access

Admin-role holders (Daniel, Rich, Tim, admin)

Watch out: It holds 2 sample projects ('3251 Toledo permitting and framing', 'Fall marketing push') and 6 tasks, untouched since Sept 7. Tim's real schedule and delay tracking happen in his Drive delay system, not here.

What it is

Gibraltar's subcontractors and vendors: 149 active, 85 verified by Tim, 101 with an email address. Sub-tabs: Directory, Campaigns (email blasts to trades or realtors, with automatic follow-ups), Formats (email templates), History and Referrals (the program where an agent owes Gibraltar 25% on a lead Gibraltar gave them).

Why we built it

To reach trades and outside agents from one list, with emails that confirm who is insured and who has actually been paid.

How it works

The list started from the QuickBooks vendor summary and was cleaned against Tim's own vendor sheet. Campaigns would send through Google mail. With sending off, every test send goes to a test address and is recorded.

How you use it

Dashboard > Trade Directory.

Who has access

Admin-role holders (Daniel, Rich, Tim, admin). It can also be granted to a project manager.

Watch out: No campaign has ever been sent, and 0 referrals are recorded. Before going live: turn sending on, connect a reply mailbox so a reply stops the follow-ups, and confirm about 30 enriched owner emails. Realtor campaigns also need an unsubscribe line and a postal address before any blast.

What it is

Runs the program for selling finished spec homes. Stage 9 sends a warning. Ticking 'blue-tape ready' arms the home. Pressing Liquidate starts the clock: 4% to the buyer's agent plus $5,000 flex cash, 3.5% on day 14, 3% on day 30, with what the agent gives up passed to the buyer as flex cash, and a review every 14 days until sold. Sub-tabs: Clocks, Notices, Intake line, Replies, Offers, Reviews and Settings (admins only).

Why we built it

The program Jim, Rich and Daniel set up for Oct 1, to move finished inventory faster with clear, pre-agreed steps.

How it works

Stage changes on Homes & Stages drive the timeline. Every notice is a draft until an admin approves it, and even then it goes only to Anthony's test inbox until two switches are turned on: one in the tab and one on the server. Neither is on. The Liquidation phone line rings Natalie, Daniel, Rich, Johanna and Jim together for 15 seconds, then hands the caller to an AI receptionist, which logs the call here.

How you use it

Dashboard > Liquidation. Liquidate a home from its Homes & Stages card.

Who has access

Admin-role holders only (Daniel, Rich, Tim, admin). Natalie, who is meant to work it, has no dashboard login.

Watch out: Today: 6 homes are in 'warned' status and none has been liquidated. All 15 warning drafts were discarded as out of date. The intake line has taken 3 calls, all tests from Sept 25 to 30. Going live waits on Rich (the listing agreement and modification templates with Jennifer Legg), on Jim (broker and license details), on Natalie (the MLS field order) and on Anthony (reading one rehearsal send, then flipping the switches).

What it is

Tracks emails Anthony sent to staff and is waiting on. It nudges people in the same email thread, and alerts Anthony the moment someone replies.

Why we built it

Anthony asked for 'automated reply encouragement' and an agent that waits on replies so nothing stalls silently.

How it works

Every 10 minutes the server checks the watched threads in anthony@'s mailbox (read-only). After 2 business days of silence it sends a nudge in the same thread (weekdays 8 AM to 6 PM ET), at most 3 per thread. The first reply marks the thread as answered and emails Anthony an alert. Separately, every time Anthony messages Claude on the Studio, it checks these threads; that per-message check has been snoozed until Mon Oct 19, 7 AM.

How you use it

Dashboard > Liquidation > Replies, or nothing: it runs by itself.

Who has access

Admin-role holders

Watch out: 4 threads are still waiting, all already at the 3-reminder maximum, so no more reminders will go out: Rich x2 (which of the 6 liquidation automation switches to turn on, Oct 1; install Discord, Oct 5), Jim (install Discord, Oct 5) and Johanna (install Discord, Oct 5). At 5:51 PM on Oct 16 Anthony closed four stale threads (the Madden stage-10 date, which has passed; the BuilderTrend stage-photo how-to to Tim; the Sep 25 liquidation overview to Rich; and the ring-list confirmation to Daniel, which Daniel answered in more depth outside the thread), and at 6:07 PM Jacob's Discord thread (he joined Oct 7). On Mon Oct 19 at 7 AM a Studio job emails Rich, Jim, Johanna, Natalie, Kelsey and Alessandra a Discord invite and closes the three old Discord threads, which leaves only Rich's automation-switch question.

What it is

A read-only view of the billing@ mailbox that sorts vendor email into Respond, Drafts, Set aside, We answered, Categories and Browse, so the person answering bills can see what is waiting on Gibraltar.

Why we built it

Vendors were waiting weeks for answers. The full-year scan found about 1,900 threads waiting on a person, the oldest 364 days.

How it works

The server follows billing@'s changes every 2 minutes (the last sync was a minute before this check) and runs a full pass every morning after 5 AM ET. It covers only the past 30 days. Kelsey can tick an email or a group off; a tick only greys the row, and it comes back if the vendor writes again. Nothing is written to the mailbox, no AI reads the mail, and nothing is drafted or sent.

How you use it

Dashboard > AP Comms. Tick items off as they are handled.

Who has access

Kelsey (full) and admin-role holders

Watch out: About 150 threads from the last 30 days sit in Respond. On Oct 16 Anthony marked the AP Comms project 100% complete; the AI sorting, mailbox labels and reply drafting items were closed as dropped from scope, not built. This tab is separate from invoice forwarding: it still shows billing@ mail, but invoices are not reaching QuickBooks (see the Vendor-invoice pipeline entry). Kelsey used it today (13 ticks).

What it is

Gibraltar's own e-signature tool. Upload a PDF or paste text, place signature fields, and send from bids@ with a private link. The signer agrees to sign electronically, enters a one-time code sent by email, and signs by drawing or typing. Everyone then gets a sealed PDF with a certificate page and tamper-evident fingerprints.

Why we built it

Anthony asked for an in-house signing tab. Liquidation's day 14 and 30 contract changes are meant to be signed here.

How it works

Signer links can be used only once and they expire, and every step is recorded in an event log. Until the tab's Live switch and a server setting are both on, every email goes to a test inbox. A full rehearsal was signed on Sept 27.

How you use it

Dashboard > E-signatures > New.

Who has access

Kelsey (full) and admin-role holders

Watch out: Only the Sept 27 rehearsal document exists. Going live is Anthony's call. A census item still asks which e-signature provider the listing changes should use, so the team should confirm this tab is the answer.

What it is

One page that counts homes in build, Gibraltar's lots, building permits, lot opportunities and FD Funding money, with a dot map of Gibraltar-owned or about-to-be-owned property.

Why we built it

Anthony asked for 'all the lots, all of the homes currently building ... buildable lot opportunities ... a map with the dots', plus permits and live FD funding.

How it works

It reuses the same numbers as the Homes, Lots and FD tabs rather than recalculating them. Permits come from Drive permit cards that the Studio pushes daily (61 today). Gibraltar-titled parcels are checked with the county every 6 hours (38 today). Money only shows to people who also hold FD Funding.

How you use it

Dashboard > Metrics.

Who has access

Admin-role holders (Daniel, Rich, Tim, admin)

Watch out: Other lenders' balances (Lima One, Renovo, MV51, CAMSA), Palm Bay permits and CRM deals are not included. About 30 small layout fixes are pending. About a third of homes have no map pin, because the lot has no street number yet.

What it is

Every Gibraltar project and task from the master to-do census (see the Hively Organizer entry), shown as a Kanban board and as a grouped list: 39 projects and 257 items, 137 done (53%).

Why we built it

Anthony's Oct 16 ask: 'quantify exactly every project in kanban style ... only rich and jim and daniel get access.'

How it works

Reads the census file live on the server and refreshes every 60 seconds. Internal notes are never shown. Groups: Systems, Liquidation program, and Waiting on people. To change what leadership sees here, the census itself has to be edited.

How you use it

ops.gibraltarbuilders.info/progress: summary on top, projects below.

Who has access

Exactly three named people, read-only: Daniel, Rich and Jim. No one else can be granted it, not even admins or Anthony's shared admin login.

Watch out: Rich was emailed about the new tab on Oct 16 (the summary is at the top; scroll down to see the projects). Jim cannot open it until he accepts his new invite (good until Oct 23). Some items Anthony closed on Oct 16 were closed by dropping the remaining work, not by finishing it (AP Comms sorting, labels and drafting; two FD sheet addresses). A few open items are already out of date (the Closings Log fix and Daniel's daily email, both repaired Oct 16).

What it is

A per-project punch list of what is done and what is still owed, kept inside Anthony's Hively platform. The 'Gibraltar Builders' project is the master to-do list for all of this work.

Why we built it

Anthony needed one honest record of done versus still-owed. On Oct 16 he reopened 12 items marked done too early and added 81 items across 11 new areas.

How it works

The list is a file on the same cloud server as the ops dashboard. Anthony edits it in the Hively dashboard's Organizer or with a helper script on the Studio. The Progress tab reads it directly. As of the evening of Oct 16: 39 projects and 257 items, 137 done (53%) and 120 open. That day Anthony marked three projects 100% complete: Mobile app, FD Funding ledger, and AP Comms (billing@ triage); their remaining sub-items (two duplicate addresses on the FD sheet; AP Comms AI sorting, labels and drafting) were closed as dropped from scope, not built.

How you use it

Read it through ops > Progress. Changes: ask Anthony.

Who has access

Editing: Anthony (Hively admin). Reading: Daniel, Rich and Jim through Progress.

Watch out: Item titles are written for Anthony. Some 'Ask X' items are open questions for staff that nobody has been asked directly yet. Some open items are already out of date: 'fix the Closings Log sync' and 'daily email to Daniel stopped Oct 15' (both repaired Oct 16), and 'open the bot to staff' (the bot already answers everyone).

What it is

A placeholder page reading 'Coming online', meant for target agents, county data sources and machine settings.

Why we built it

Reserved for central settings.

How it works

No features yet. Real settings live inside each tab: Liquidation > Settings, E-signatures > Settings, and the Superintendents panel in Homes & Stages.

How you use it

Nothing to do here yet.

Who has access

Alessandra, Daniel and the other admin-role holders

Watch out: The Homes card says to add stage-checklist items 'in Settings', but no screen exists for that yet; only Anthony can add them.

What it is

The admin screen for dashboard logins: create people, email invites or set temporary passwords, give each person full or view-only access tab by tab, mark someone as a project manager or superintendent (Homes only, their own homes), make someone view-only everywhere, switch a login off, and resend an invite as a password reset. The same screen also manages Sales CRM tab grants for people who have a CRM account.

Why we built it

So access can be handed out tab by tab, for example Kelsey sending e-sign documents while only reading FD, without giving anyone more than they need.

How it works

The server is the real gatekeeper. Hiding a sidebar link is only cosmetic; data from a tab you do not hold is refused, and view-only tabs reject every change. One role exists, Administrator, which sees every tab except Progress. The system refuses to remove or switch off the last admin. Switching a login off ends its sessions and kills any unused invite. Progress-style named-people tabs cannot be granted at all.

How you use it

Dashboard > Users & Access (admins only).

Who has access

Administrators: Daniel, Rich, Tim and the shared admin login.

Watch out: Tim's login holds the Administrator role, so he can see FD Funding money, P&L and this screen, even though his own tab list only names stages, homes, the timeline, lots and trades. Natalie and Johanna have no dashboard login.

What it is

A ChatGPT-style chat, plus hands-free voice (tap the orb and talk), that answers questions from live company data: homes and stages, FD Funding, sheet sync status, and the Drive knowledge base.

Why we built it

To let staff ask questions in plain English instead of digging through tabs and Drive.

How it works

The page passes your question through Cloudflare to a service on the Mac Studio. The service checks your own dashboard session every time and looks up data with your own access. Thinking is done by Groq (an outside AI service). Speech-to-text uses Groq Whisper, and the spoken replies use ElevenLabs. There are daily limits on voice and chat. The conversation disappears when the tab closes.

How you use it

Dashboard > Assistant (top of the sidebar).

Who has access

Everyone signed in to the dashboard. Admins get the broader company-data tier; others get the staff tier.

Watch out: Only Anthony's admin login has used it (17 chats between Oct 2 and Oct 5), and nothing since. Questions and the data used to answer them go to an outside AI provider. Undecided data-exposure findings: the staff tier of the company knowledge base can include signed contracts and escrow documents.

What it is

The sales team's CRM: Leads, Inbox (conversations with buyers), Pipeline, Tasks, Calendar, Marketing, Campaigns, Import leads, Team and CRM Settings. It holds 1,021 contacts. Inside ops it appears under a 'Sales CRM' sidebar heading at /sales/...; the phone-first version is at ops.gibraltarbuilders.info/crm/ (and crm.gibraltarbuilders.info).

Why we built it

So sales and operations share one website and one sign-in, while agents keep a phone-friendly CRM to use in the field.

How it works

A separate Cloudflare app with its own database. Meta ad leads arrive through Zapier, and new leads go to agents in rotation, each told immediately. Sales calls are logged from the phone system. The ops sidebar shows only the CRM sections your CRM account allows. CRM permissions live in the CRM, not in the dashboard. Most CRM changes come from Daniel's own Claude on his MacBook Pro (about 134 entries in the shared activity log, against about 49 from Anthony's Studio). Daniel's seat also maintains the sales-line ring order; today it shipped a contact-metrics fix and an 'owner first' ring change. Anthony's ops dashboard hosts the CRM at /crm.

How you use it

Dashboard sidebar > Sales CRM, or ops.gibraltarbuilders.info/crm/ on a phone.

Who has access

11 CRM accounts. Admins: Daniel, Natalie, Rich and a shared Admin. Agents: Gretza, Jennifer, Jodie, Kris, Steven, Randy and Julie (Julie's is switched off). Gretza, Jennifer, Jodie, Kris and Steven are in the lead rotation.

Watch out: Outgoing CRM messages send within a minute of being queued, with no approval step; only quiet hours and opt-outs stop them. That contradicts the 'a person approves every send' design and is still undecided. CRM texts have gone out for real through Twilio since Oct 6 (3 delivered through Oct 9; earlier ones were test stubs), under another business's texting approval. 5 phones have push notifications on. Johanna has no CRM account.

What it is

A Light or Dark switch at the bottom of the dashboard sidebar.

Why we built it

Dark was hard to read in daylight and in print; light was asked for alongside it.

How it works

Your choice is remembered per browser. Dark is the default. Printing always uses light. Email previews, the buyer portal, and labels on solid colored fills deliberately ignore the switch.

How you use it

Sidebar footer > Light or Dark.

Who has access

Everyone signed in

Watch out: In dark mode the Trade Directory still has a lot of low-contrast grey text; light mode fixes it.

What it is

The dashboard can be installed like an app on Android, iPhone or desktop, named 'Gibraltar Ops', so it opens full-screen from the home screen and stays signed in.

Why we built it

Superintendents and the office approve from their phones, and a sign-in in a normal browser tab is forgotten when the tab closes.

How it works

The site tells the browser how to install it. The first sign-in screens show a Chrome 'Install' button plus written steps for Android, iPhone and desktop. Once installed, the sign-in is remembered. If the connection drops, the app keeps you signed in and retries instead of logging you out.

How you use it

In Chrome: menu > Install app. On iPhone Safari: Share > Add to Home Screen.

Who has access

Everyone with a login

What it is

A native iPhone/iPad app called 'Gibraltar' (App Store Connect name 'Gibraltar Ops') that wraps the live dashboard. It adds phone-friendly touches: pull to refresh, PDFs and downloads that open signed in, printing to the share sheet, an offline screen, and automatic reload after each update.

Why we built it

Anthony: 'build the full mobile app, make everything mobile compatible ... so we can ship this thing.'

How it works

The app loads ops.gibraltarbuilders.info, so every website update reaches phones without a new app build. No Mac here has Xcode (Apple's app-building software), so builds are made and signed in the cloud by Codemagic under Anthony's Apple developer team (the same team that signs Hively's apps). It is distributed through internal TestFlight to the group 'Gibraltar staff'.

How you use it

Ask Anthony to add you as a tester. Staff must first be added as users on his Apple team.

Who has access

Testers: only Anthony today. Tim, Kody and Jacob have not been invited.

Watch out: TestFlight builds expire after 90 days, so this build lapses around Dec 28 and needs a rebuild by about Nov 28 (only Anthony can start one). Push notifications are supported by the app but the server side is not built. A bad website update breaks the app for everyone at once.

What it is

Two finance mail robots that run inside the ops server. The invoice robot reads the billing@ mailbox, picks out real vendor invoices and forwards them to QuickBooks Online's receipt-capture inbox so they can be posted to the books. The closing-document robot reads kelsey@ for closing statements and lot-closing packages and puts them on the FD Funding Pending list for Kelsey.

Why we built it

To stop forwarding invoices and closing paperwork by hand. Invoices went to Dext until Aug 14 and have gone straight to QuickBooks's receipt capture since then.

How it works

Every 10 minutes the ops server opens billing@ and sorts each new email in a fixed order. First a loop guard, then a reminder veto, so a payment reminder for a bill that was already filed is never entered twice. Next an AI classifier on Groq, and keywords only if the AI gives no answer. Real invoices go into a send queue. Every decision is logged. The cloud server cannot send mail itself, so a small program on the Mac Studio (still called the 'Dext send agent') emptied that queue every 5 minutes and did the actual forwarding. The kelsey@ robot works the same way but only saves the PDFs into the dashboard; it sends nothing.

How you use it

Normally it runs by itself, but right now nothing reaches QuickBooks. Until Anthony repairs it, invoices sent to billing@ have to be forwarded by hand. The AP Comms tab still shows billing@ mail because it reads the mailbox through a different Google connection. Closing documents still show up on the FD Funding Pending list.

Who has access

The robots log in to billing@ and kelsey@ with stored mailbox credentials. Anthony maintains them, and Kelsey works the results.

Watch out: Checked live today. The Studio sender stopped on Aug 26. Its schedule line is commented out with a note that says 'paused Aug 14, invoice capture moved off Dext', but its log shows it kept delivering to QuickBooks until Aug 26 at 2:05 PM: 98 invoices reached QuickBooks between Aug 14 and Aug 26. 126 invoices queued between Aug 26 and Sep 15 have never been sent. The billing@ reader has failed every run since Sep 15 (its mailbox password stopped working), so mail after Sep 15 is not even being sorted. The census already lists 'move the 126 queued vendor invoices' as open. The kelsey@ robot is working: it last checked at 5:55 PM today, connected. Still unverified: whether QuickBooks splits an email with several attachments into separate bills.

What it is

The GB Inventory Sheet, the Google Sheet where Sales keeps every home and lot with its price, status and dates, is the shared source behind several of Anthony's systems. A few of them also write back to it.

Why we built it

So there is one list of homes. Ops, FD Funding, Homes & Stages, Metrics, the Closings Log and the connectors all work from the same rows Sales already maintains.

How it works

Every 10 minutes the ops server reads the sheet and refreshes its own copy of each home (last check today 5:55 PM, in sync). In the same pass it reads the sheet's 2026 Production tab for build milestone dates; today 43 of 56 rows matched a home and 13 addresses did not. Some FD Funding edits are written back to the sheet through a Google script bridge. The Closings Log on the Studio fills only EMPTY cells from settlement statements and never overwrites anything. The Studio connector's 'home status' tool reads the sheet live. Since Oct 16, Tim's Drive relay is also allowed to write to it.

How you use it

Change prices and statuses in the sheet itself; ops picks the change up within about 10 minutes. If a page warns about 'sheet drift', or a home goes missing, a column was probably moved or renamed: tell Anthony.

Who has access

Sales edits the sheet. The ops server and the Studio relays read it, and some of them write as described above.

Watch out: The sheet lives in admin@'s own Drive, not the company Shared Drive. The census lists an open decision because the sheet is shared by link. Moving or renaming columns can silently break the readers. 13 Production-tab rows do not match any home.

What it is

A private practice copy of the Homes & Stages system that runs only on the Mac Studio. It uses a copy of real data with every email and text sender switched off, so people can be shown the system or trained on it without any buyer getting anything.

Why we built it

Built on Sep 25, 2026 so Anthony could walk Tim through the stage system live without touching a real buyer.

How it works

It copies the previous night's backup of the ops database, cuts the link to the Google Sheet, and adds throwaway admin and project-manager logins on test addresses. It then starts the app on spare ports on the Studio, with every mail and text setting blanked. The Studio's real robots only talk to production, so they never touch the copy. To demo photos, the photo processor is run by hand against the copy. Every click was measured to record what it writes.

How you use it

Ask Anthony to start it for a demo or training session. It opens in the browser panel on the Studio.

Who has access

Anthony, on the Studio

Watch out: The data is only as fresh as the backup it was copied from. Stage checklists are empty in production too, so they have to be filled in by hand for a demo. Homes with real buyers are deliberately avoided as demo homes.

What it is

This is Anthony's main AI workstation: Claude Code running on the office Mac Studio under the machine's "rich" login. Every other piece in this section either feeds it or is reached from it.

Why we built it

Anthony does all of Gibraltar's dev and AI work from this machine. Keeping one hub means one set of notes, one set of connections and one place to apply safety rules.

How it works

A Claude session here can read files on the Studio's disk, including the raw copy of the company Drive. It can run commands, reach the ops server in the cloud by hopping through the Mac Mini, and use the connectors listed below. Claude's own permission rules and an automatic safety check apply on top, and that check has refused several risky requests, such as loosening the document-tier rules or giving the Discord bot admin powers. Every shell command a Claude session runs is appended to a local log. Persistent notes ("memory") record what was built and why.

How you use it

Ask Anthony. Daniel has sometimes driven Anthony's Studio session remotely through Claude's Remote Control feature.

Who has access

Anthony at the keyboard, plus anyone he lets in through Remote Control (Daniel has used this).

Watch out: This session is NOT read-only. It can run commands, change files and send email as anthony@. When Daniel uses Remote Control, he is driving this full session, not the read-only connector described next.

What it is

One read-only connector on the Mac Studio that every Gibraltar AI uses to look up company information. It answers questions from the Drive documents, the live GB Inventory Sheet, the Closings Log and the Gibraltar Systems Guide.

Why we built it

Anthony wanted one shared "executor" so every AI agent gets the same company abilities instead of each being wired up separately. On Oct 13 he extended it for Daniel: "connect Daniel Casas' computer to the Mac Studio MCP so he can see the closings."

How it works

It offers six tools. kb_search searches all company Drive documents by meaning. kb_browse walks the Drive folder tree. read_document returns the full text of one document. home_status reads the live GB Inventory Sheet. closings reads the Closings Log, one row per settlement statement, and only leadership-level keys get it. systems_guide returns the internal systems guide. Every caller sends its own key, and the connector filters what comes back by that key's level (next entry). It listens only on the Studio itself and on Tailscale (a private, encrypted network that links the company's Macs), never on the office Wi-Fi or the open internet. Every request, tool call and failed login is logged. There have been about 1,600 tool calls since Oct 2, most of them automated tests.

How you use it

There is nothing to open. You use it by asking your own Claude a Gibraltar question (see "Daniel's Claude connection"). Anthony creates and revokes keys on the Studio.

Who has access

Holders of a connector key: Daniel's Claude, Tim's Claude, Claude on the Studio and on the Mac Mini, and the ops Assistant. The Discord bot has a key that is not wired in yet.

Watch out: Read-only: no tool can change a Drive file, a sheet or anything else. Searches take 15 to 45 seconds when the Studio is busy. It works only while the Studio is on and the caller is on Tailscale. Scanned PDFs with no text layer can't be found (for example, about 27 scanned closing PDFs as of Oct 13). The closings tool is only as fresh as the Closings Log sync. That sync timed out on every run from Oct 15 at 6:45 AM, then ran cleanly at 3:12 PM, 3:32 PM and 4:05 PM on Oct 16 (186 closings). All returned text is labelled as data, not instructions, so a document can't steer the AI.

What it is

Each AI that uses the connector has its own named key. Each key has one of three levels: staff, dev (leadership) or owner. Separately, every company document is sorted into one of four sensitivity tiers, and the level decides which tiers a key can see.

Why we built it

So that a bot anyone can message sees only what any staff member could, leadership AIs see leadership material, and the most sensitive files go only where Anthony chose owner level.

How it works

The four tiers: - T0, open: marketing, realtor-facing and homeowner documents. - T1, internal: most project files. - T2, restricted: Finance, Office Opps and corporate files, contracts, buyer files, closing and money documents, and the GB Inventory Sheet. - T3, highest: HR, payroll, wires and escrow, bank and card authorisations, W-9/1099/SSN, and anything mentioning passwords or logins. What each level gets: - Staff: searches T0 and T1, reads only T0 and the Systems Guide in full, sees home status with dollar amounts and dates stripped out, and has no closings tool. - Dev (leadership): T0 to T2, the leadership home columns, and closings. - Owner: everything including T3, unredacted, plus every column of the inventory sheet's Sales and Production tabs. The company login documents are blocked for every level, owner included. Keys sit in each Mac's Keychain, and the connector stores only a fingerprint of each key, not the key itself.

How you use it

Ask Anthony to add, change or revoke a key. A change takes effect immediately, with nothing for the user to re-run.

Who has access

Keys by level (no key values shown): - studio-claude (dev): Claude on the Studio. - mini-claude (dev): Claude on the Mac Mini. - daniel-claude (owner since Oct 13). - tim-claude (owner since Oct 16; it was staff when created Oct 9). - ops-assistant-dev (dev): used when an ops admin uses the Assistant. - ops-assistant-staff (staff): used for everyone else on the Assistant. - hermes (staff): created for the Discord bot but not connected. - test-staff and test-dev: test keys only.

Watch out: Tiers are decided mostly by folder and file name. A sensitive file saved in an ordinary folder under an ordinary name can land in a lower tier, and this is an open finding: signed contracts and escrow items sit in T1. Claude sessions running on the Studio itself can open the raw Drive copy on disk, so the tiers limit the connector, not Anthony's own sessions. On Oct 16 Tim's key was raised to owner on Anthony's instruction, so Tim's Claude now sees the same HR, payroll and wire documents that Daniel's does.

What it is

Daniel's own Claude, on his Mac Studio, is connected to the knowledge connector with an owner-level key. It can answer from company documents, the live inventory sheet and the Closings Log.

Why we built it

Anthony, Oct 13: "connect Daniel Casas' computer to the Mac Studio MCP so he can see the closings." Then: "leadership access across all tools... nothing he shouldn't see; goal = run reports on closings."

How it works

A one-time setup file, emailed to Daniel from anthony@ on Oct 13, did four things: - stored his key in his Mac's Keychain; - registered the "gibraltar" connector for Claude Code; - installed a small local bridge so the Claude desktop app's Chat and Cowork can use it too; - added a short instruction telling his Claude to check the connector first for anything about Gibraltar and to name the documents it used. His Mac and the Studio see each other through a Tailscale share. Anthony chose a connector-only link, with no remote login in either direction. It first connected Oct 13 at 5:43 PM. That evening it made 25 tool calls: home status, searches, document reads and one closings report. No calls have been logged since.

How you use it

1) Turn Tailscale on. 2) Open Claude on your Mac (Claude Code or the Claude desktop app). 3) Ask in plain words, for example: - "List every 2026 home-sale closing with price, lender and title company." - "Where does 1740 Madden stand?" - "Find our warranty claim process." Adding "use the gibraltar closings tool" helps for closing reports. 4) If Claude says it can't reach the connector, check that Tailscale is on and the office Studio is awake. To remove it: run bash ~/.local/share/gibraltar-kb/gibraltar-kb-setup.sh --undo, or ask Anthony to revoke the key.

Who has access

Daniel (key daniel-claude, owner level)

Watch out: It works only from Daniel's Mac with Tailscale on. Claude on the web or on a phone can't reach it, because claude.ai's remote connectors can't reach the private network. The built-in instruction names four tools and doesn't mention closings, so ask for closings by name. Closing answers are only as good as the statements the title companies emailed in, and rows marked "review" need a person to confirm. The inventory sheet's Production "closing date" column is mostly empty, so most closing facts come from the Sales tab's free-text status.

What it is

Tim Mayes' Claude, on his MacBook, uses the same connector with the same installer as Daniel's. Since Oct 16 its key is at owner level.

Why we built it

Anthony, Oct 9: "set him up with the knowledge connector and make sure all his Claude chats going forward use it." On Oct 16 he gave Tim owner-level access on every tool.

How it works

Same design as Daniel's: the key is in his Keychain, the connector is registered for Claude Code, a local bridge serves the desktop app, and an instruction block tells his Claude to check the connector first. He ran the setup on Oct 12. The connector has logged 30 calls from him, the latest Oct 16 at 3:13 PM, now at owner level.

How you use it

Tim asks his own Claude. Nothing is needed from Daniel.

Who has access

Tim (key tim-claude, owner level since Oct 16)

Watch out: Owner level includes the T3 documents (HR, payroll, wires and escrow). Only the login sheets are withheld.

What it is

Every 15 minutes the Mac Mini copies, read-only, Tim's Claude notes, his Claude workspace documents and his Claude and Cowork chat histories. The office Claude can then search what Tim has been working on.

Why we built it

Anthony, Oct 9: "make sure tims computer is sharing its claude knowledge base with the mac studio." Later: "we just need all access ... index everything."

How it works

A locked-down key on Tim's MacBook lets the Mini copy four folders and nothing else: it can't delete or run anything. His private Finance and Personal Finance chat folders, and anything his own ignore-lists exclude, are left out. The Mini indexes his notes and workspace documents (107 today), and the Studio's Claude searches them through the hively-kb connector. A summariser was built to turn raw chat transcripts into short session records using only the Studio's local model. It produced 17 records on Oct 12 and 13 but has no schedule, so it isn't running.

How you use it

Nothing for Daniel. Claude on the Studio searches it when a question touches Tim's work.

Who has access

Claude on the Studio and Anthony's own devices. Tim's laptop is not itself allowed to search it.

Watch out: Tim's Mac must be awake and online for the copy to run. Raw transcripts are stored on the Mini but not indexed. The summariser's privacy filter let 16 of 35 test personal-finance items through, so switching it on is held for Anthony's decision. Studio sessions working "as Tim" (in ~/tim) take notes copied from Tim's Mac as instructions, a known risk Anthony accepted.

What it is

A locked-down channel that lets Tim's Claude hand questions and whole jobs to the office Studio's Claude, using a "gb-studio" command on his Mac.

Why we built it

Anthony, Oct 12: let the two Claudes talk, so the Studio does the heavy reading and building and Tim's own usage allowance lasts.

How it works

There are two lanes. - "ask" gets an answer from a read-only Studio Claude. It can read the ops code, the delay-system index and Tim's workspace, and has no shell, web or connectors. - "request" files a job. A dispatcher starts a separate worker Claude for each request, at most three at once. Since Oct 13 Tim's request counts as the approval. A safety hook on every worker refuses moving money, entering or storing credentials, permanent deletes, deploying Anthony's own systems (ops app, CRM), editing the bridge itself and reading secret files. Refused steps are skipped and listed in the reply, never queued for Anthony. Tim can watch progress live. The nightly Delay Log pull (12:45 AM) also runs as a bridge request.

How you use it

Tim uses gb-studio from his Mac. Nothing is needed from Daniel.

Who has access

Tim's MacBook only, through one pinned key

Watch out: Workers CAN make live changes. They write to Drive and Sheets through the relay (as rich@), to JobTread through Rich's connector, and to Tim's mailbox, and they can read every company inbox. To switch the bridge off, Anthony removes one line from the Studio's list of allowed keys.

What it is

A complete local copy of the Gibraltar Builders shared Drive on the Studio, about 10,600 files and 14 GB. It is turned into searchable text and indexed by a model that runs on the Studio.

Why we built it

Rich asked for the AI to be "trained" on the Drive without spending tokens. Anthony built search instead of training: search stays current every night, cites real files and can't invent figures.

How it works

Each night at 2:30 AM, rclone (a file-copy tool) copies the Drive using a read-only Google permission, so it can never write back. Text is pulled out of each file, scanned PDFs are read with local OCR (text recognition), and a local model indexes about 315,000 passages across about 5,650 documents. A separate "live" job checks Drive about every 5 minutes and indexes only changed files, so new documents appear within minutes. Each night the system also publishes a tiered "snapshot" for the connector and the Discord bot, and redraws the "brain" picture (a network map of the documents that the bot can show). Claude chat transcripts from the Studio are ingested too, for local search only, never through the connector.

How you use it

Through any connected Claude (yours, the ops Assistant, the Discord bot). On the Studio itself: - kb "question" returns matching passages with Drive links in under a second. - ask "question" has the local model write a cited answer in 60 to 90 seconds, free, and nothing leaves the Mac.

Who has access

The connector's key holders (filtered by level) and Claude sessions on the Studio

Watch out: Since Oct 14 the nightly publish has refused to run. Its safety check stops it because the leadership-tier document count jumped more than 20% (from 1,058 to 1,362). Until someone confirms the jump and re-runs it, the connector and bot use the Oct 13 snapshot plus the minute-by-minute changes on top. The nightly runs on Oct 14 and Oct 16 took about 29 and 13 hours during the runaway-process incident, Oct 15's run never started, and Oct 11's Drive copy failed. While the nightly runs, the live refresh waits, so on Oct 16 no Drive changes reached search between 2:39 AM and about 3:23 PM. Load is back to normal. Two files fail text extraction every night. A search score under 0.50 means "not in the Drive". Older scanned PDFs without OCR are invisible (an Archive backlog of about 1,325 was noted in August). The Drive copy still uses rclone's shared Google sign-in client, which Google is retiring during 2026.

What it is

A 63-page plain-English guide to every internal system: the ops dashboard tabs and their addresses, Homes & Stages, the CRM, FD Funding, AP Comms, the sales site, Discord and the bot, who does what, a glossary and quick links.

Why we built it

On Oct 1 the Discord bot didn't know the ops dashboard address. Anthony said it "needs more knowledge".

How it works

An AI workflow (several readers, then a writer, then a checker) wrote it on Oct 1 from the code and the notes, with no internal addresses. It is stored on the Studio as its own knowledge source (it isn't in the Drive). The systems_guide tool serves it, and it also appears in search results.

How you use it

Ask any connected Claude, e.g. "Use the systems guide: how does AP Comms work?"

Who has access

Every connector key. Staff and dev levels see 58 of the 63 pages; owner keys see all 63.

Watch out: It was written Oct 1 and hasn't been updated, so it misses everything built since: the ops Assistant, Metrics, P&L, the Lots revamp, the Closings Log, the Progress tab and the connector itself. One page is wrong, not just old: 'Vendor invoices from billing@ (where they go)' says invoices are forwarded to QuickBooks about every 10 minutes and that this kept working through the move to QuickBooks Online. In fact nothing has been delivered since Aug 26 and billing@ has not been read since Sep 15. Hermes, the Assistant and the connector all answer from this guide, so they will repeat the error until the page is corrected. Five sign-in pages are hidden from staff and dev keys because they contain the word "password".

What it is

A second read-only connector, on Anthony's Mac Mini, for searching the AI working notes ("memory") kept on Anthony's machines and the notes shared from Tim's Claude.

Why we built it

One place where every device can look up what other AI sessions already learned, instead of rediscovering it.

How it works

It identifies callers by their Tailscale login instead of a key. The office Studio gets a narrow role: it can search only its own notes (80 documents) and Tim's shared Claude work (107). Anthony's own devices see everything, including a pass-through to the Gibraltar connector. Everyone else is refused, including Tim's laptop. The index refreshes every 15 minutes and anything shaped like a key or password is blanked out.

How you use it

Claude on the Studio uses it automatically. Daniel doesn't use it directly.

Who has access

Claude on the Studio (narrow role) and Anthony's own devices (full)

Watch out: It also holds notes about Anthony's other businesses, which are deliberately hidden from the Studio. The Studio's own notes, which hold a lot of Gibraltar detail, are copied to the Mini every 15 minutes, so they are readable from Anthony's devices.

What it is

Two-way, password-free secure login (SSH) between the Studio and Anthony's Mac Mini. The Mini is also the stepping stone the Studio uses to reach the ops server in the cloud, where the ops dashboard, the Gmail module and the Drive relay run.

Why we built it

The Mini holds the fullest record of Anthony's projects, and it is the machine with access to the ops server and to GitHub.

How it works

The Studio logs into the Mini with its own dedicated key, and the Mini then logs into the ops server. This route carries: - the per-message email check; - sending email as anthony@; - the Drive and Sheets relay; - Tim's mail tool; - read-only database checks; - calls to Astra. GitHub account work (issues, CI runs) is meant to run on the Mini as AnthonyRijos, the organisation owner.

How you use it

Nothing for Daniel. Anthony and Claude use it.

Who has access

Claude sessions on the Studio and Anthony

Watch out: The Mini's GitHub login has been invalid since Sep 28 (re-checked today), and only Anthony can sign it back in. Ordinary code pushes from the Studio still work through a repository key. If the Mini is down, the email check, sending and the Sheets relay all stop.

What it is

AI on the Studio reads and sends company email through the ops server's Gmail module, which holds a Google Workspace permission over company mailboxes.

Why we built it

Anthony, Sep 17: any email the system sends goes out as anthony@, "the voice the business speaks to vendors and buyers in." The Resend email service isn't usable because none of its accounts has a verified sending domain.

How it works

Google's permission covers every company mailbox (read, send, label), and the code narrows it. Sending and labelling are allowed only as billing@, scheduling@, kelsey@, anthony@, bids@ and, since Oct 16, tim@. Since Oct 16 every other @gibraltarbuilders.com mailbox can be READ but never sent from. Anthony added this so Tim's Studio requests can read "emails across the org" through the tim-mail tool. Delete permission was never granted, so nothing in the system can destroy mail. The service key lives in one locked file on the ops server.

How you use it

Ask Anthony. Claude sessions, reply tracking and Tim's workers use it.

Who has access

Claude sessions on the Studio and Tim's bridge workers (through tim-mail)

Watch out: Since Oct 16, leadership mailboxes, Daniel's included, can be read by Tim's Studio requests. The running ops dashboard picks up the tim@ send change only at its next restart, but the tim-mail tool already uses it.

What it is

A command that lets Claude on the Studio, and Tim's request workers, read and write Google Sheets, Docs and Drive as rich@, through the company's Google service account on the ops server.

Why we built it

Built Oct 9 for Tim's weekend delay-tracking run, then widened on Oct 16 when Tim was given owner-level access.

How it works

Each request goes through the Mini to a fixed script on the ops server, which enforces the rules. There are no deletes, no trashing, no sharing or permission changes, and no deleting of sheet tabs it didn't create. Delay Log writes are refused from 11:40 PM to 12:40 AM, a quiet window just before the 12:45 AM nightly pull. Since Oct 16 every file in the Gibraltar Shared Drive, plus the GB Inventory Sheet, is writable through it. Before that, it could write only to Tim's runbook files and files it had created.

How you use it

Nothing for Daniel. Tim's requests and Anthony's sessions use it.

Who has access

Claude sessions on the Studio and Tim's bridge workers

Watch out: Edits show as rich@ in Drive history. Anthony's standing rule for the GB Inventory Sheet: AI may fill empty cells only, never overwrite or delete an existing value.

What it is

The "Assistant" tab on the ops dashboard: a chat and voice assistant that answers from live company data.

Why we built it

To let the team ask plain questions (a home's status, FD Funding balances, where a document is) inside ops instead of hunting through tabs.

How it works

Your browser talks to a service on the Studio through a Cloudflare tunnel (a secure pipe from the internet to the Studio with no open ports), and the service re-checks your ops login on every request. The model is an open model hosted by Groq (gpt-oss-120b), not Claude. Groq's Whisper turns speech into text, and ElevenLabs' "Sarah" voice reads replies aloud. Its tools: - the knowledge connector (admins get the leadership key, everyone else the staff key); - FD Funding, Homes & Stages and sheet-sync status, read through the ops dashboard with YOUR login, so it can never show more than your own tabs allow. Limits: 300 chats per person per day and 40,000 spoken characters per person per day.

How you use it

Go to ops.gibraltarbuilders.info/assistant (every signed-in user sees it), then type or tap the microphone.

Who has access

Every signed-in ops user. Admins (Daniel, Tim, Rich and the shared admin login) get leadership-level knowledge; the others get staff level.

Watch out: Not used since Oct 5: 29 chats in total, mostly Anthony's tests. Questions and tool results go to Groq and ElevenLabs. Anthony has no named ops login, only the shared admin one. Groq's tier is small and shared with other ops jobs, so it falls back to smaller models when busy. If the Studio is off, the tab can't answer.

What it is

The company Discord bot answers from the same Drive knowledge base and live inventory sheet at staff level, using an AI model that runs on the Studio itself.

Why we built it

Anthony chose Discord as company comms on Sep 30, with the bot as the way staff get answers.

How it works

Its own add-on reads the knowledge-base snapshot and the inventory sheet directly; it isn't on the connector yet. Its model (qwen3.6) runs on the Studio, so text questions don't leave the building. Voice notes go out through Groq (speech to text) and ElevenLabs (speech). It can also control the Studio's screen for Anthony only (see Hermes desktop control): background clicking, typing, scrolling and app switching were set to 'always allow' on Oct 1, and screenshots may be posted in any channel.

How you use it

@mention Hermes in the company Discord or DM it. Today the server's people are Daniel, Anthony, Tim and Jacob; Rich, Jim, Johanna, Natalie, Kelsey and Alessandra get invites on Mon Oct 19 at 7 AM.

Who has access

Everyone in the Discord server (since Oct 8): today Daniel, Anthony, Tim and Jacob. Desktop control is Anthony's alone.

Watch out: The bot's connector key was created but never wired in. The #admin channel is readable by everyone in the server, and staff have now joined. Whether company email and Discord chat should go into its knowledge has been held for Anthony's decision since Sep 30. The census item 'Open the bot to staff' is out of date: the live setting already answers everyone.

What it is

A skill is a packaged instruction sheet that Claude loads by itself when a task matches it, like a runbook. Two Gibraltar skills are installed: gibraltar-kb and fd-funding. A set of general claude.ai skills (Word, Excel, PDF, PowerPoint, Google Docs and others) is also installed.

Why we built it

So every Claude session on the Studio handles Gibraltar work the same way without being re-taught each time.

How it works

gibraltar-kb tells Claude to search the Drive copy with kb or ask before reading whole folders, and to say "not in the Drive" instead of stretching a weak match. fd-funding is the operator runbook for the FD Funding system: the dashboard, its background agents, the two-way Google Sheet bridge and the related scheduled jobs.

How you use it

Nothing to do. Claude picks the skill up when a question is about Drive documents or FD Funding.

Who has access

Claude sessions on the Studio

Watch out: The fd-funding skill is stale. It was last edited Jul 31 and still describes the old setup: the dashboard at developer-digital.cc (that address no longer loads), the retired shared-key login, Netlify deploys, the QuickBooks Desktop relay (dead since the move to QuickBooks Online) and the Dext forwarder. A Claude session that loads it for 'check the invoices' gets wrong instructions until it is rewritten.

What it is

Small scripts that run every time Anthony sends Claude a message on the Studio and add up-to-date context before Claude answers.

Why we built it

Anthony, Oct 4: "every time I type in chat make sure to read my email to check if we have responses." The bridge and watchdog lines were added Oct 12 and Oct 16.

How it works

Three lines are added to each message: 1) Email reply check: read-only. It checks the tracked reply threads and the last 24 hours of team mail to anthony@, at most once every 90 seconds, through the Mini to the ops server. 2) Bridge summary: new or updated requests from Tim's Claude, plus warnings. 3) Watchdog line: the Studio watchdog's alert summary, with a warning if the watchdog itself has stopped. A separate read-only Claude (the "Studio Watcher") reviews each new alert.

How you use it

Nothing for Daniel. To switch the email check back on early, Anthony deletes the snooze file.

Who has access

Anthony's sessions (the checks are skipped in Tim's worker sessions)

Watch out: The email check was snoozed at 5:45 PM on Oct 16 until Mon Oct 19 at 7 AM ET because everyone is out until Monday; it switches back on by itself. The bridge and watchdog lines still run, and the watchdog shows no alerts. Only 4 tracked threads are still waiting (two to Rich, one to Jim, one to Johanna), and each has already had the maximum 3 reminders, so no more reminders go out. The command log has grown to 33 MB.

What it is

claude.ai's JobTread connector, signed in as Rich Pierce. Claude on the Studio and Tim's request workers use it for JobTread work such as delay tracking, task dates and links.

Why we built it

Tim's delay-tracking system and the Oct 9 to 12 weekend run needed JobTread. JobTread can be reached only through this connector, not from scheduled scripts.

How it works

JobTread's own sign-in grant, named "Access for claude.ai", was created Oct 9 at 10:23 PM ET and expires Jan 14, 2027. Checked today: it allows reading, creating, updating AND deleting across JobTread, covering jobs, tasks, daily logs, documents, vendor bills, payments and customers. Anthony replaced an earlier read-only grant with this one for the weekend run.

How you use it

Ask Anthony, or Tim through the bridge. Changes show in JobTread as Rich Pierce.

Who has access

Claude sessions on the Studio signed into the claude.ai account that holds the grant, which includes Tim's bridge workers

Watch out: Earlier notes call it read-only; that is no longer true. Claude's own rules and the worker safety hook refuse permanent deletes and money moves, but the grant itself would allow them. Tim's own JobTread grant lives in his claude.ai account and can't be used from the Studio.

What it is

claude.ai's Google Drive connector, attached to Claude sessions on the Studio. It can search, read, create, copy, update, share and trash Drive files.

Why we built it

Lets Claude find or produce a Google file directly when the knowledge base or the relay isn't the right fit.

How it works

It is a separate path from the read-only Drive copy and from the Sheets relay. Most Gibraltar document reading goes through the knowledge base, and most sheet writing through the relay.

How you use it

Ask Anthony.

Who has access

Claude sessions on the Studio

Watch out: Which Google account it is signed into, and how often it is used, were not checked. It is capable of sharing and trashing files, so Claude's rules require an explicit OK for those.

What it is

An AI video, image and voice generation service connected to Claude sessions on the Studio through Anthony's paid account.

Why we built it

Used for marketing and explainer material: the liquidation demo video (Sep 25) and two 60-second explainer videos for the central comms proposal. It was also checked for 3D home tours, and it can't make walkable tours.

How it works

Claude picks a visual style, generates scenes and narration, and stitches the video together locally. No company data is needed beyond the script.

How you use it

Ask Anthony.

Who has access

Claude sessions on the Studio (Anthony's account)

Watch out: It costs credits (about 281 for the two explainers). Its output should be used only as clearly labelled marketing extras, never as real photos of homes.

What it is

Tools that let Claude drive a web browser or the Studio's screen when a system has no proper data connection.

Why we built it

Some systems have no usable API (a programming interface other software can call), BuilderTrend's photo pages for example.

How it works

Examples: mapping BuilderTrend's photo pages once, after which a fixed robot does the repeated clicking with no AI involved; filming the liquidation demo video; working in web portals. The Discord bot's desktop control (above) is a separate, Anthony-only version of the same idea.

How you use it

Ask Anthony.

Who has access

Claude sessions on the Studio

Watch out: It acts as whoever is signed in to the Studio's Chrome. Claude's rules forbid typing passwords and making purchases, and sending anything needs an explicit OK.

What it is

Four open AI models that run on the Studio itself: embeddinggemma (builds the search index), qwen3.6 (the Discord bot's brain and the ask command), gemma4 31B and qwen3-vl 30B (which can read images).

Why we built it

Free, unlimited, and nothing leaves the building, which suits company documents.

How it works

Ollama serves the models on the Studio only. The knowledge base, the ask command, the Discord bot's text answers and Tim's session summariser all use them.

How you use it

Nothing to do directly. They sit behind the tools above.

Who has access

Services on the Studio

Watch out: Heavy jobs compete for the Studio's processor. During the Oct 13 to 16 runaway-process incident, searches slowed and nightly runs took many hours.

What it is

Two outside AI companies that receive some company data today. Groq runs the ops Assistant's model, voice transcription and the ops dashboard's email and bank-line classifiers. ElevenLabs provides the spoken voice for the Assistant and the bot, and runs the Liquidation phone agent.

Why we built it

They are cheap and fast for jobs the local models can't do in real time (live chat, live voice).

How it works

The FD Funding email watcher sends Groq the sender, subject, attachment names and the start of each email from billing@ and kelsey@, including closing statements. That text can contain buyer names, addresses, loan numbers and payoff amounts. Bank transaction lines are classified the same way.

How you use it

Nothing to do. These run inside the ops dashboard and the bot.

Who has access

Ops server and Studio services

Watch out: Sending closing-document details to Groq was never formally decided. Nobody has reviewed ElevenLabs' retention and training terms for recorded calls.

What it is

OpenAI's gpt-6-astra model, run through OpenAI's Codex command-line app on Anthony's Mac Mini. Claude calls it as an independent reviewer from a different AI company.

Why we built it

A second opinion from another vendor catches mistakes Claude might share with itself. Sep 20 review: Astra plus a fresh Claude session fill the "independent critic" role better than Grok.

How it works

Claude on the Studio sends Astra a prompt over the Mini link in a read-only sandbox, and Astra writes a review file. It was used on the central comms decision (it agreed with extending gb-crm, on condition that someone is named to maintain it) and on the 3D home tours plan.

How you use it

Ask Anthony.

Who has access

Claude on the Studio and Anthony

Watch out: It reviews only what it is sent and never touches company systems. Which OpenAI account pays for it isn't recorded in the notes.

What it is

The Hermes command-line agent on the Studio, signed in to Rich's PERSONAL Grok subscription (model grok-4.5). There is no company xAI account or API key.

Why we built it

It was used as an independent critic in a handful of reviews. It found real defects, and was confidently wrong twice.

How it works

After a spending-limit block on Sep 29, Grok calls work again: the usage records show successful calls on most days from Oct 1 to Oct 16. About 85% of its use is a background search-ranking worker that isn't Gibraltar work. In August and September it was used a handful of times as an outside reviewer of Gibraltar code and plans, so that code went to xAI under a personal consumer login. A Gibraltar route called grokask was built with strict limits: Marketing and Production documents only, with no contracts, deeds, pricing, Sales, Finance or HR. It has never run, because it needs a company API key that doesn't exist.

How you use it

Not used for Gibraltar work now. Ask Anthony.

Who has access

Anthony

Watch out: The Sep 20 review said to keep it out of anything that touches customers. xAI's usage rules ban automated housing and credit decisions, and its consumer terms give xAI broad rights over whatever is sent, which includes the Gibraltar code it reviewed. A planned one-hour comparison of Grok, Claude and Astra was never run.

What it is

A profit forecast, done with Claude on the Studio, for the 12 listed or under-contract homes due to finish by Dec 31.

Why we built it

Anthony asked on Oct 16: 'project the profit on the next few houses we have currently listed, looking historically and at the prices and calculator'.

How it works

It uses Finance's own definition of net proceeds, the same one the P&L 2026 tab uses for sold homes: price minus lot, build cost, build interest, other job costs, commission, seller credit, survey and taxes, and lender fees at closing. Inputs, read from the Drive knowledge base copy: Finance's funding workbook (each home's lender and loan terms), the Jul 8 cost-per-project report from QuickBooks, the May 28 master budget per plan (cost to complete), and the sales contracts and closing statements. It assumes each home sells 5.9% under list, the 2026 average. A first pass built on plan averages came out about $12k per house too optimistic, so it was rebuilt and checked.

How you use it

Ask Anthony for the write-up. Refreshing it needs a newer cost-per-project report or the QuickBooks Online connection, which is not built yet.

Who has access

Anthony; the results were reported in his Claude session

Watch out: Base case: about $17,500 in total for the 12 homes before overhead (overhead runs about $44k per house sold in 2026). Range: -$228k to +$278k. 6 of the 12 lose money. 1740 Madden closed Oct 13 at about $49k profit, which Finance has not booked yet. An $80,000 note on Amberwood was not modeled. FD draw fees are booked at company level, not per house. None of this is in ops.

What it is

A chat assistant called "Gibraltar Hermes" in the company Discord. You @mention it in a channel or send it a direct message, and it answers questions about the company: where documents are, how our systems work, and where a home stands.

Why we built it

Anthony decided on Sep 30 that Discord would be the company chat and Hermes the bot in it. On Oct 1 he said the bot "needs more knowledge" (it didn't know the ops web address), so a plain-English guide to every system was added to what it knows.

How it works

It runs on the Mac Studio on a free AI model that runs on the Studio itself (Qwen 3.6), so ordinary questions never leave the building. Its tools: (1) a search of the company Google Drive mirror, limited to staff-level documents. Finance, contracts, HR, payroll and the company login sheets can never be searched. (2) A 60-plus page "Gibraltar Systems Guide" written for it in plain English. (3) A live lookup of a home on the GB Inventory Sheet. It shows address, plan, stage, estimated completion and agent, and never price, lenders or contract dates. (4) A /brain command that posts a picture of what it knows. It may only post links to ops, crm, sales and the public website. Since Oct 8 it answers anyone in the server. Before that it answered Anthony alone.

How you use it

In the Gibraltar Builders Discord, @mention Hermes in any channel or DM it. Type /new to start a fresh conversation and /brain for the knowledge picture.

Who has access

Anyone in the Gibraltar Builders Discord server (since Oct 8). Today that is Daniel, Anthony, Tim and Jacob. Desktop control and the permit-portal check are Anthony's alone.

Watch out: Because its AI model is a small one, it makes mistakes. In an Oct 8 test it could not finish a dashboard-access request and mislabelled the dashboard, so check its answers. So far only Anthony has used it: 19 Discord messages, all his, the last on Oct 7-8. Five Systems Guide pages, the sign-in instructions, are hidden because they contain the word "password". No channel is set up as a private leadership room, so everyone, leaders included, gets staff-level answers. It is not yet connected to the Gibraltar executor that Daniel's Claude uses (a staff key exists but is not wired in). Discord's rules ban financial or sensitive data, so FD Funding and closing documents are kept out. This is not the older "Grok" Hermes seat, which runs on Rich's personal xAI login and is not used for company chat.

What it is

Lets Anthony ask the bot, from Discord, to take a screenshot of the Mac Studio's screen and to click and type on it. A second owner-only tool reads the status of the robots that check the permit portals.

Why we built it

Anthony asked on Sep 30 for it to "use the full computer... on questions it can't find the answers to... send picture updates".

How it works

Only messages from Anthony's main Discord account can start it. His second account ("Ant") cannot; making it an owner was blocked on Oct 8. Screenshots need no approval, and the bot reads them with the same AI model. Every click or keystroke was designed to post an Approve button that only Anthony can press. On Oct 1, however, he chose "always allow" for background clicking, typing, scrolling and switching apps, so those steps no longer stop for approval. Screenshots may be posted in any channel; Anthony said everyone in the Discord is an admin. The permit-portal tool only reads and never clicks.

How you use it

Anthony only, from his main Discord account. No one else can use it.

Who has access

Anthony's main Discord account only

Watch out: Screenshots of the Studio desktop can land in channels every member reads. With "always allow" on, the approval guard on desktop control is thinner. One setting switches it off.

What it is

The bot can reply with a spoken voice note, and it understands voice notes people send it.

Why we built it

On Oct 7 Anthony asked that "send me the response as a voice note" always work, and that the bot should "reason what I'm asking". It now answers by voice when someone seems to want to listen, for example while driving or on site.

How it works

Spoken replies use the ElevenLabs voice "Sarah" through the key for Daniel's ElevenLabs workspace. Incoming notes are transcribed by Groq Whisper. Limits: 1,200 characters per person per day (Anthony is exempt), 4,000 a day and 40,000 a month company-wide, and 5 minutes per incoming note. Links and file paths are removed before speaking, and it refuses anything that looks like a password or key.

How you use it

DM the bot or reply to one of its messages. Either send a voice note or ask "send that as a voice note". In channels, a voice note without an @mention is ignored.

Who has access

Everyone in the Discord server

Watch out: Used once so far: one 6-second transcription on Oct 7. The ElevenLabs credits are shared with the Liquidation line's AI agent, and the key we hold cannot read the balance. A single setting switches voice off.

What it is

Ask the bot in Discord to text or email someone. It drafts the message and posts a card with Send and Cancel buttons. Nothing goes out until the person who asked taps Send.

Why we built it

Anthony asked for this on Oct 5: texts and emails "with a confirm button afterward". He also decided that anyone, customers included, can be texted or emailed, starting that day.

How it works

The bot only writes drafts. Tapping Send calls the ops server, which texts through Twilio from the company's toll-free main line (the 833 number) and emails through Google as anthony@. Rules: only the asker's tap counts, not even Anthony's; a tap in the first 4 seconds is refused; each draft sends once; texts to people outside the company only go out between 8 AM and 8 PM ET; daily limits apply. A 13-name staff list lets "text Jim" find the right person.

How you use it

In Discord, @mention or DM Hermes, for example "text Natalie that..." or "email Kelsey about...". Read the card, then tap Send.

Who has access

Anyone in the Discord server (since Oct 8): today Daniel, Anthony, Tim and Jacob. Rich, Jim, Johanna, Natalie, Kelsey and Alessandra get invites on Mon Oct 19 at 7 AM.

Watch out: Older notes call this "not live", but it is. The ops server starts in LIVE mode, 3 real texts went out on Oct 5 along with 1 test email, and nothing has been sent since. On the first day a mistyped number made the bot text Jim by mistake. The same-day fix makes the bot ask before guessing who a typed number belongs to. Texts go out under another business's texting approval (see Business texting registration), so carriers could block them. Since Oct 8 any server member can send to anyone. The ops server keeps a record of each send.

What it is

A private Discord server, "Gibraltar Builders", with channels #general, #production, #sales, #marketing, #finance and #admin and a voice room. Hermes lives in it.

Why we built it

Anthony made Discord the company chat on Sep 30. September research had compared Mattermost, Google Chat, Slack and Discord, and Discord was the free, simple option.

How it works

Daniel's account owns the server, and Anthony uses it to manage settings. Anthony sent invitations by email from anthony@ on Oct 5 (Rich, Jim, Johanna, Tim) and Oct 7 (Jacob); Tim and Jacob joined. Anthony approved fresh invites: a Studio job emails Rich, Jim, Johanna, Natalie, Kelsey and Alessandra the #general invite link from anthony@ at 7 AM on Mon Oct 19 and closes their old reply-watch threads. It sends once, and only between Oct 19 and Oct 23. Anthony also asked whether Claude should create Discord accounts for staff from their company mailboxes; that was declined in favour of invites, because Discord accounts are personal (each needs its own password and phone) and Discord's terms expect people to make their own.

How you use it

Install Discord, open the invite link from Anthony's email, then @mention Hermes with questions.

Who has access

Members today (6): Daniel (owner), Anthony (two accounts), Tim, Jacob and the Hermes bot. Not in the server: Rich, Jim, Johanna, Natalie, Kelsey, Alessandra and Kody.

Watch out: Still to decide: whether to invite Kody (Natalie, Kelsey and Alessandra get Monday's invites). Anthony decided to raise the verification level, require 2-step sign-in for moderators and switch to an expiring invite only after everyone has joined; until then the invite link never expires and anyone who gets it can join. The September research ruled Discord out as a place to keep records: bots cannot read staff DMs, and there is no export or retention. Use it for chat, not records. When last checked (Sep 30 to Oct 1), the server did not require two-factor sign-in for moderators and every channel, #admin included, was readable by every member. That has not been re-checked, and staff (Tim, Jacob) are now in the server.

What it is

How every Gibraltar system sends and reads email. Messages go through the company's own Google Workspace as real mailboxes, so they land in the real Sent folder and thread normally in the recipient's inbox.

Why we built it

Rich, Sep 17: "anytime i ask to email make sure it from anthony@". That way mail reads as the company, not the owner chasing something personally. Resend, the outside email service used before, refused every Gibraltar address because no domain had been verified with it.

How it works

A Google service key with company-wide permission lives only on the ops server. The code lets it send as six mailboxes: anthony@ (the default), scheduling@ (buyer stage updates), bids@ (trade directory and e-signatures), billing@ (vendor invoices), kelsey@ (closing documents) and tim@ (added Oct 16). Since Oct 16 it can also read, but never send from, any other company inbox, for Tim's requests. It has no permission to delete mail. Test sends go to Anthony's personal test inbox. Google's email signature record (DKIM) is now published for the domain.

How you use it

Nothing day to day. When you ask Anthony, or Claude on the Studio, to send something, it goes out from anthony@.

Who has access

Anthony runs it. The key itself can reach every mailbox in the company.

Watch out: The list of six mailboxes in the code is the only limit on a key that could read or send for anyone, so treat it as the company's most sensitive secret. Resend is effectively dead: it still has no verified domain, yet the Studio's buyer stage-update sender is still set to use it (stage updates are on hold anyway), and the CRM's automated noreply email depends on it. The domain's DMARC policy is monitor-only, so spoofed mail is not rejected.

What it is

A watcher that remembers the emails where we are waiting for someone's answer. It reminds them politely in the same thread and tells Anthony the moment they reply.

Why we built it

Anthony asked on Sep 27 for "automated reply encouragement... then an agent waiting on a reply so we know".

How it works

Every 10 minutes the ops server checks each watched thread in Gmail. With no reply after 2 business days, it sends a short reminder in the same thread from anthony@ (weekdays, 8 AM to 6 PM), up to 3 times. The first reply marks the thread replied and emails Anthony. A separate check on the Studio looked at these threads every time Anthony typed to Claude. Anthony snoozed that check on Oct 16 at 5:45 PM until Monday Oct 19 at 7 AM, because everyone is out until Monday; it switches back on by itself. 4 threads are still waiting, all already at the 3-reminder maximum, so no more reminders will go out: Rich x2 (which of the 6 liquidation automation switches to turn on, Oct 1; install Discord, Oct 5), Jim (install Discord, Oct 5) and Johanna (install Discord, Oct 5). At 5:51 PM on Oct 16 Anthony closed four stale threads (the Madden stage-10 date, which has passed; the BuilderTrend stage-photo how-to to Tim; the Sep 25 liquidation overview to Rich; and the ring-list confirmation to Daniel, which Daniel answered in more depth outside the thread), and at 6:07 PM Jacob's Discord thread (he joined Oct 7). On Mon Oct 19 at 7 AM a Studio job emails Rich, Jim, Johanna, Natalie, Kelsey and Alessandra a Discord invite and closes the three old Discord threads, which leaves only Rich's automation-switch question.

How you use it

Open ops.gibraltarbuilders.info, then Liquidation, then Replies, to see who we are waiting on. To answer, just reply in the same email thread.

Who has access

Ops admins see the Replies tab. Alerts go to Anthony.

Watch out: Reply alerts still go to an old personal Gmail that Anthony asked on Oct 5 to stop using. Reminders only go to staff, never customers.

What it is

The system behind the plan to sell finished spec homes faster. When a home reaches stage 10 ("blue-tape ready"), it can be liquidated: better commission for the buyer's agent, flex cash for the buyer, a running clock, and the paperwork and notices for the team.

Why we built it

From the Sep 22 sales meeting (Jim, Rich, Daniel, Natalie, Johanna and Anthony, with John Curry on the phone). On Sep 25 Anthony asked to "start the liquidation trigger engine, campaign machine, AI intake line".

How it works

It watches stage changes in Homes & Stages. At stage 9 it drafts an early warning to Daniel, Rich, Natalie, Johanna and the listing agent. At stage 10 someone presses Liquidate. Terms follow Daniel's Oct 5 rulings: 3% to the buyer's agent before liquidation; 4% plus $5,000 flex cash at Liquidate; 3.5% on day 14 and 3% on day 30, with price drops of $5k and $10k; the commission the agent loses goes to the buyer as flex cash; never 5%; no end date, with a review every 14 days until the home sells. It produces the three listing-agreement change sheets, and the signed PDF is uploaded on the home's card. Sub-tabs: Clocks, Notices, Intake line, Replies, Offers, Reviews and Settings. Every outgoing message is a draft until an admin approves it. Approved messages still go only to Anthony's test inbox unless BOTH the tab's Live switch and the server's switch are on. Once live, signing reminders (the day before and the day of) go out on their own, to company addresses only. Six send switches are all OFF: program notices, intake team alerts, the follow-up text to callers, email blasts to outside agents, the CRM special-offer broadcast, and live e-signatures.

How you use it

ops.gibraltarbuilders.info/liquidation (admins). Each home's card in Homes & Stages also has a Liquidation button.

Who has access

Ops admins: Daniel, Rich, Tim and Anthony (the shared admin login). Natalie has no ops login yet. Project managers can only tick blue-tape and read.

Watch out: Today: 6 homes are in the warned stage, none have been liquidated, and no notices are waiting (15 out-of-date warnings were thrown out Oct 4-5). Nothing has been sent for real. The main blockers are people: the listing agreement and change templates (Rich, with Jennifer Legg), Natalie's MLS access (Daniel), and which switches to turn on (Rich). The automatic MLS listing write-up and listing pages are not built. Some meeting ideas were deliberately left out: the one-at-a-time ring order (the team agreed to ring everyone at once instead), the $250 assist bonus and the lowball rule.

What it is

A dedicated phone line for buyers and agents calling about liquidation homes. A call rings five leaders' cell phones at once. If nobody takes it within 15 seconds, an AI receptionist answers and records the caller's details.

Why we built it

The meeting's rule was "no voicemail": every caller reaches a person or the AI. Daniel decided on Sep 29 that these calls are handled in-house, because sales agents would want commission.

How it works

The public number is a RingCentral line (user "Liquidation Line") set to forward every call to a hidden Twilio number bought for this purpose. Twilio asks the ops server what to do, and it rings Natalie, Daniel, Rich, Johanna and Jim together for 15 seconds. Whoever answers hears "Gibraltar showing line. Press 1 to take this call", so a voicemail box cannot grab the call. If nobody presses 1, the call goes to the AI. Each call becomes one row in Liquidation, then Intake line.

How you use it

When it rings, answer and press 1. Calls and the AI's notes appear at ops, Liquidation, Intake line, where admins can also change the ring list and ring time.

Who has access

Ring list: Natalie, Daniel, Rich, Johanna, Jim. Settings are in the Liquidation tab (admins).

Watch out: Only three calls ever, all tests on Sep 25 and Sep 30, and none since. The Monday Oct 5 ring test was missed (Natalie's reply: the 9 AM agents meeting ran over) and has not been re-run. The number is not published yet, because notices are off. Weakness of ringing everyone at once: a single phone that sends calls straight to voicemail (off, Do Not Disturb, Silence Unknown Callers) can end the ringing for everyone, and the caller goes to the AI. Team alert emails and texts are off and have no recipients. No "this call may be recorded" notice plays before the AI answers, even though the AI keeps transcripts and Florida requires every party's consent. A spare RingCentral call queue made on Sep 30 is unused.

What it is

The AI voice that answers the Liquidation line when nobody picks up: "Thanks for calling the Gibraltar showing line. Which property are you calling about?"

Why we built it

So no liquidation caller is ever lost to voicemail.

How it works

Built on Sep 29 in Daniel's ElevenLabs workspace (Creator plan), using GPT-4o and the "Jessica" voice. It asks whether the caller is an agent with a buyer or a buyer, which property, when they want to see it, their agent and brokerage, and their move timeframe. It checks the property against our listings, saves the answers to the ops server, says a fixed goodbye and hangs up, with a 5-minute limit. Transcripts are pulled from ElevenLabs when someone opens the call.

How you use it

Nothing: it answers on its own. Read and work the calls in Liquidation, then Intake line, marking each scheduled, referred, junk or done.

Who has access

Daniel owns the ElevenLabs workspace, and Anthony has a seat in it.

Watch out: Proven by one real test call on Sep 30. Its ElevenLabs key is overdue to be replaced (it was exposed on Sep 30; see the open questions). Hively's demo copy of the agent and its borrowed toll-free number still exist and should be removed. The agent saves answers through a Hively web address, because the ops server currently runs on Anthony's Hively server. Answers are not yet written onto the CRM contact.

What it is

Every step of the liquidation program as a checklist. 93 tasks were filed on Sep 22 and are now part of the company's project list.

Why we built it

Rich asked on Sep 22 for "a developer task tree... broken down step by step - e.g. if I need to get credentials for this to finish that".

How it works

Each task names its owner, its steps, what it needs, what blocks it and where a human must approve. Tasks are grouped by who does them: Anthony's builds 0 to 8, logins needed, "Waiting on Jim/Daniel/Natalie/Johanna/Rich", and a social auto-poster on hold. They are stored in the Close Out List of the Hively Organizer and shown on the ops Progress tab. As of Oct 16: 25 of 71 build tasks are done; 4 of the 28 tasks waiting on people are done. Still open: Jim 10, mostly decisions (launch homes, blue-tape standard, call recording, weekend coverage, lender links, Zillow/homes.com, referral-fee routing, flex-cash figure and video, his agent contacts); Daniel 5 (Natalie's MLS access, signing path, MLS listing fields, brokerage social collaborations and blast rules, flex-cash script); Natalie 5 (listing-agent roster, agent contacts sheet, shooting, editing and publishing the flex-cash video); Johanna 2 (closing statements into Drive, her agent contacts); Rich 1 (the listing agreement, which is the first domino).

How you use it

ops.gibraltarbuilders.info/progress (Rich, Jim and Daniel only)

Who has access

Rich, Jim and Daniel can view Progress. Anthony edits the list.

Watch out: Some tasks were ticked done without anything being built, and the Oct 16 audit reopened them; the MLS listing write-up is one. Task notes are limited to 2,000 characters, so some steps were cut short.

What it is

A directory of subcontractors and suppliers (149 active, including 14 realtors and brokers) with email campaigns and automatic two-day follow-ups. The same tool is meant to email liquidation offers to outside buyer's agents.

Why we built it

Built Sep 16 to reach trades by email. Extended Sep 25 as the liquidation "campaign machine" that emails outside agents when a home is liquidated.

How it works

The trades came from the QuickBooks vendor summary and Tim's vendor directory (85 verified by Tim). Campaigns send from bids@ through Google. While sending is switched off, every message goes to a test inbox and is logged. There is no reply mailbox connected, so replies are marked by hand.

How you use it

ops.gibraltarbuilders.info/trades (admins and Tim)

Who has access

Ops admins and Tim

Watch out: No campaign has ever been sent: 0 campaigns, 0 sends. 48 active trades have no email, including 31 that Tim verified, and 30 emails come from a paid skip-trace and are unconfirmed. Realtor emails have no unsubscribe link or postal address, which the law requires before any blast. About 70 outside-agent emails from past deals are ready, but are waiting on Natalie's shared agent sheet, and Jim and Johanna were asked to add their own contacts.

What it is

Tracks referral fees when an agent gets a free lead from us (the meeting set 25% internal and 30% external), and checks them against closing statements.

Why we built it

Build 6 of the liquidation program. It extends Realtor Accountability, which already asks which lead led to which sale.

How it works

Referrals are logged with a tier and later matched to closings harvested from Drive. It sends nothing.

How you use it

ops, Trade Directory, Referrals tab

Who has access

Ops admins and Tim

Watch out: No referrals logged yet. Who gets paid (through Curry to Daniel, or through Developer Digital) is undecided and waits on Jim. Harvesting closing statements waits on Johanna's folder.

What it is

The map of Gibraltar's phone numbers by purpose, and which of them can send texts.

Why we built it

Checked on Oct 12, when Anthony asked whether the normal RingCentral number can text.

How it works

On RingCentral, Gibraltar's phone system: the company main line cannot text. The Sales Team queue line can. The Liquidation line can, but only from that line's own app login, which Anthony holds. Tim's line cannot, and Anthony's own app login has no texting number. Across the whole RingCentral account, 0 texts were sent in the 180 days before Oct 12. On Twilio (Daniel's "Thoughts In Form" account): RingCentral department numbers forward calls into the toll-free main line (the 833 number). It rings one sales rep at a time for 20 seconds, then the ElevenLabs "Sales Call Receptionist" answers. The same line is what the CRM, the Discord bot and the planned caller follow-ups text from. Twilio also holds the hidden Liquidation bridge number, which is voice only. Measured Sep 17 to Sep 30 on the sales line: 25 calls, 8 answered by a person, 12 by the AI, 5 hung up.

How you use it

To text, use the RingCentral app on the Sales Team or Liquidation line. Letting the main number text means adding it to RingCentral Business SMS, which needs Anthony's RingCentral password and an OK.

Who has access

RingCentral admins (Tim is a Super Admin; Anthony reaches the Admin Portal through his app login). Daniel owns the Twilio account.

Watch out: The Twilio balance was about $15 on Oct 5, with spending around $22 a month. At $0, the main line's calls AND texts stop. A RingCentral forward hides which department number the caller dialed, which is why the Liquidation line got its own hidden number. The sales line's code now says "This call may be recorded" and records calls a rep answers; whether that version is deployed was not checked.

What it is

The carrier approval a business needs to send texts. The company's toll-free main line texts today under an approval that was filed for another business. Gibraltar's own registration was started on Oct 5 and has not moved since.

Why we built it

Carriers filter or block unregistered business texts, and the CRM, the Discord bot and the planned liquidation follow-ups all text from that line.

How it works

Checked Oct 5: the toll-free number is ours, but its texting approval was filed in April for a different business, not for Gibraltar. An older registration attempt failed in March. Replies to texts from that number probably reach nobody. Anthony accepted the risk on Oct 5. The new registration is for the hidden Liquidation number: the "GIBRALTAR BUILDERS LLC" business profile is submitted and in review, and the brand ($4.50, low volume) is pending. The website now has an optional text-consent checkbox on /contact and a text messaging section in the privacy policy.

How you use it

Ask Anthony. Daniel owns the account and is the named representative.

Who has access

Daniel (account owner and named representative); Anthony works the console

Watch out: The next step, filing the campaign (about $15 to vet plus a monthly fee), needs a Twilio top-up and Anthony's OK. Nothing has happened since Oct 5. Adding the hidden number to a texting service must not break the live Liquidation call routing. The CRM ignores the website's text-consent checkbox and does not store it. If carriers revoke that other business's approval, every text from the main line fails. The address on the IRS record may not match the Sarno Rd office, which could get the brand flagged.

What it is

The Sales CRM that Daniel runs (crm.gibraltarbuilders.info) sends texts and emails to clients from a queue that empties every minute, and no person approves each message. This is the plain flag.

Why we built it

The every-minute timer exists so new web leads get an instant reply. It was flagged on Sep 20 because the September comms proposal to Rich promised "a person approves every send". That is true of the design but not of the code.

How it works

Every minute the CRM takes any queued outgoing message and sends it. The only checks are quiet hours and opt-outs; there is no approval field anywhere. Texts go through Twilio from the toll-free main line. One-to-one emails go from the agent's own Gmail when they have connected it (since Oct 12). Automated email uses a noreply address through Resend. Checked live Oct 16: the CRM's live-sending switch is ON (turned on in early October). So far 3 real texts were delivered (Oct 6 and Oct 9, one an automatic appointment reminder) and 1 email (Oct 12). Nothing is queued right now, and no drip sequences or automation rules are active (one sequence sits in draft), so today only appointment reminders and messages people send by hand go out.

How you use it

Daniel and the sales agents use the CRM. Ask Daniel.

Who has access

Daniel (admin) and the sales agents

Watch out: The moment anyone activates a sequence or broadcast, messages to buyers go out with nobody approving them. The September research recommended adding an approval step before connecting any new channel. The CRM ignores the website's text consent. Its AI receptionist talks to callers on its own and stores transcripts. Its texts go out under another business's texting approval.

What it is

A planned new brand for $1M to $5M custom homes in Brevard.

Why we built it

Daniel's recap of the Oct 7 Customs meeting.

How it works

The plan: fixed price at a 30% margin, a 10%-of-profit bonus for the project manager, and 3% sales commission at closing. Leads come from free outreach first (lot owners, listing agents, our agents), with paid ads later. Onboarding: intake form, response call, then an onboarding call with the designer and project manager (1.5% design deposit). Owners: Daniel runs Meta ads and the landing page. Anthony owns the Google Business Profile, Google ads and the cold-outreach pipeline. Tim and Jacob handle lead response. Kelsey handles owner financing. Rich holds a meeting on the bonus standards. On Oct 7 Anthony asked Tim for past custom homes to turn into AI-written SEO portfolio pages.

How you use it

Ask Daniel or Anthony.

Who has access

Daniel, Anthony, Tim, Jacob

Watch out: Daniel, Oct 12: the launch is on hold per Jim and Tim, but the Google Business Profile should still be built and the custom-home information still collected. No Customs site, profile or files were found. The lots revamp plans a search for Customs lots from about December to February.

What it is

The public website. Daniel rebuilt it on Sep 24 as hand-written pages hosted on Cloudflare Pages, replacing WordPress, and Anthony added search-engine (SEO) work on top.

Why we built it

Before the work, on Sep 15, Gibraltar appeared in Google results for 0 of 25 tracked Brevard home-builder searches and in none of the AI answers.

How it works

The code lives in the company GitHub repo, and a push to the main branch goes live in about 30 seconds. Anthony's SEO script adds to every page: business data that tells Google and AI tools we are a home builder, page descriptions, social preview images, a sitemap, an llms.txt file for AI tools and FAQ data. City pages show county tax-roll facts. New sections: a Build on Your Lot page (Sep 25), and Palm Bay, Merritt Island and homepage Brevard sections (Oct 6). The "no middleman" wording was removed, because outside agents sell our homes. Google Search Console was verified Sep 21, and new pages are pushed to Google and Bing. Daniel added a Home Buying Process page on Oct 12. Six of the seven model pages embed Matterport 3D tours.

How you use it

Visit gibraltarbuilders.com. Daniel edits the pages. After page edits, Anthony re-runs the SEO script.

Who has access

Daniel (builds and edits, from his own Claude on his MacBook Pro), Anthony (SEO, mostly from his Mac Mini, some from the Studio). DNS is in Cloudflare.

Watch out: Rankings have not been re-measured since the Sep 15 baseline. Most pages are still short (under 400 words). Bing, Yahoo and YellowPages still show the old address and phone. Nobody knows which account owns the main Google Business Profile, and manager access for Daniel and Natalie waits on Jim. Authorship is shared: in the shared code activity log about 26 website entries come from Daniel's MacBook Pro (the .com launch, logos, floor plans, flyers, the Home Buying Process page), about 25 from Anthony's Mac Mini (most of the SEO) and about 12 from the Studio.

What it is

Research and a first test of filming homes on an iPhone and turning the video into walk-through 3D tours for the website.

Why we built it

Asked on Sep 29: film a home on an iPhone 17 Pro, put a 3D walkthrough on the site, and see whether Higgsfield can help.

How it works

Ten research agents plus an independent review concluded: keep the Matterport tours already on six model pages, and run a one-week pilot on an empty repeat-plan home with "Gaussian splatting", which turns ordinary video into a 3D scene you can walk through. A walkthrough-video watcher has run on the Studio every 10 minutes since Sep 7: any video put in its inbox folder, and any video over 100 MB that lands in the Studio's Downloads folder (for example by AirDrop), is moved out and turned into a 3D scene automatically. It last processed test clips on Sep 8. The first full walkthrough was built on Sep 29 from two test clips on Anthony's MacBook Air. The clips were too dark.

How you use it

Ask Anthony. The research report is saved.

Who has access

Anthony

Watch out: Blocked on: who owns and pays for the Matterport account (nobody knows), Cloudflare storage (R2) not being turned on, who films (perhaps Jacob at stage 10), and a consent rule for occupied homes. Higgsfield cannot make a walkable tour, only labelled extras. Remove GPS location from clips before publishing. The project is paused, but the Studio's watcher is still running: a large video dropped into the Studio's Downloads folder will be moved out of Downloads without warning. Nothing else has changed since Sep 29.

What it is

How Anthony makes short videos for leadership: a 6 minute 51 second walkthrough of the liquidation program (Sep 25) and two 60-second explainers for the comms proposal (Sep 15).

Why we built it

To show the team how the liquidation system works before go-live, and to pitch the comms options to Rich.

How it works

He records the ops dashboard running on a local copy with all sending blanked out (a "demo sandbox"), capturing the Chrome window directly. Title cards, AI explainer scenes and narration come from Higgsfield, Anthony's paid credit-based account. The pieces are assembled on the Studio. The second versions of the explainers were built on the Mac Mini.

How you use it

Ask Anthony for a new video. The method and scripts are saved.

Who has access

Anthony

Watch out: Higgsfield credits come out of Anthony's account. Higgsfield cannot take a screen recording, so final assembly is done locally. The flex-cash explainer (Jim at a whiteboard) is a separate video for people to film, and it is not started.

What it is

A 30 to 45 second video of Jim at a whiteboard explaining how the $5,000 flex cash works, cut for YouTube, social media and agent emails.

Why we built it

Part of the liquidation marketing plan (build 7) from the Sep 22 meeting.

How it works

Daniel writes the script, Jim decides the figure he says on camera, Natalie films and edits three sizes with captions and publishes to the Gibraltar YouTube channel. Anthony then adds email, text and post templates to the CRM and the Trade Directory.

How you use it

Nothing yet. It is waiting on Daniel, Jim and Natalie.

Who has access

Daniel, Jim, Natalie, Anthony

Watch out: 0 of 2 build tasks done. Every person step is still open.

What it is

A planned feature that drafts each liquidation home's MLS listing from our data and publishes a branded page for that home on gibraltarbuilders.com.

Why we built it

Build 3 of the liquidation program.

How it works

Planned: an AI fills a standard listing form from the inventory sheet, the model pages and the plot plans, checks it against MLS rules, and renders a landing page. Natalie posts to the MLS by hand.

How you use it

Nothing yet

Who has access

n/a

Watch out: 1 of 7 tasks done. The audit found that a task marked done had never been built and reopened it. The MLS and Facebook cannot be automated (accounts get banned), and no Gibraltar branding is allowed on MLS assets. Waiting on Daniel to define the MLS fields.

What it is

A queued plan: turn a photo drop in a Drive folder into draft social posts, with captions written by the Studio's own AI model, post to the Google Business Profile through Google's official tools, and remind Natalie to post by hand where automation is banned.

Why we built it

From the Sep 22 meeting's marketing plan.

How it works

Nine tasks are written and nothing is built. Facebook and the MLS stay manual, because automated posting gets accounts banned.

How you use it

Nothing yet

Who has access

n/a

Watch out: Needs manager access for Daniel and Natalie to the Google Business Profile, Facebook, Instagram and YouTube (Jim), and a decision on LinkedIn (Daniel).

What it is

A second AI-answered phone line for Daniel, built in the style of the AI line Anthony runs for one of Hively's clients.

Why we built it

From the Sep 22 meeting, which asked for that model with "his own flow".

How it works

Planned: a second ElevenLabs agent with its own number and script, and CRM routing that tags calls by which number was dialed.

How you use it

Nothing yet. Waiting on Daniel's decisions.

Who has access

n/a

Watch out: 1 of 7 tasks done. Daniel has to decide the line's purpose, what the AI calls itself, the number, the routing and which workspace it lives in.

What it is

Tim's system for working out each house's build-time bonus. It counts calendar days from the underground plumbing inspection pass to the Final inspection pass, split at Texture: Kody Arnold owns the front half, Jacob McCreary the back half, and Tim the whole run. Excused delays are taken off before the house is graded Great, Good or Over. Excused means weather, municipal, permitting or AP, and only once Tim confirms them.

Why we built it

To pay project-manager and VP bonuses on fair, evidence-based build times instead of raw calendar days.

How it works

Everything lives in Google Drive under Production. The Delay Log sheet holds the delays, candidates, forecast and dashboard tabs. The Critical Path Registry lists which JobTread tasks count. The Delay System Control folder holds the Project Brief (the rules, which win over everything), the Job Board (J01 to J61 work items), Dispatch Rules and Job Plans. The superintendents propose delays and only Tim can confirm them. The GB Inventory Sheet's 2026 Production tab reads the delay days from the Delay Log summary. As of today the Delay Log has 26 proposed delays, 0 confirmed and 0 rejected. The Job Board shows 43 jobs verified, 10 reported but not verified, and 1 planned (J19, the how-to and announcement for Kody and Jacob).

How you use it

Open the Delay Log and Delay System Control in the shared drive (Production folder). Tim confirms or rejects delays at his Monday production meeting, from the Final Review page, once a house passes Final. Changes to the system go through Tim, either directly or by a request from his Claude to the Studio.

Who has access

Tim (owner), Kody and Jacob (propose delays), anyone with access to the Production folder for reading; Studio workers write as rich@ through the Studio's Google relay

Watch out: No delay has been confirmed yet, so no bonus has been graded with excused delays. Two Job Board items (J60, prune duplicate JobTread links; J61, sort tasks within stages) have been paused since Oct 12 waiting on Tim's choices. Tim wanted his private notes and workspace pushed to a private GitHub repo. That has not happened: he has no GitHub account.

What it is

A read-only watcher on the Studio. It keeps a plain-English summary of Tim's delay system and a running change log.

Why we built it

Anthony, Oct 9: index Tim's delay-tracking system and keep pulling its changes, so any Claude can answer questions about it without re-reading every sheet.

How it works

Every 15 minutes it reads the Studio's local copy of the shared drive (mirrored from Google Drive) and Tim's synced Claude notes. It then writes two files. INDEX.md has the current Job Board, Delay Log counts, last nightly pull and file list. CHANGES.md lists, newest first, job status moves, new or changed delay rows, nightly pull runs and Tim's note edits. It never writes to Drive or JobTread. Its last run was 5:44 PM today.

How you use it

Nothing: it runs by itself. Ask the Studio's Claude (or the read-only bridge ask lane) 'what changed in Tim's delay system' and it reads these files.

Who has access

Studio Claude sessions; Tim's Claude through the bridge ask lane (read-only)

Watch out: The watcher cannot show Tim's git history yet. The share from his Mac leaves out the git folder on purpose.

What it is

Each night it reads the superintendents' JobTread daily logs and proposes delays in the Delay Log. Tim's rule is that the Mac Studio always runs it, never his MacBook.

Why we built it

Tim's request (Oct 13): 'own the Delay Log nightly pull as a recurring scheduled task on the Studio'. His laptop missed nights when it was asleep.

How it works

At 12:45 AM a small Studio job files the night's pull as a bridge request, using Tim's own written spec. The dispatcher hands it to a worker Claude (see the workers entry). The worker reads JobTread read-only and backs up the tabs first. It turns logs where the superintendent filled in the delay fields into Proposed rows, adds keyword matches as Candidates, and writes one Pull State row. The pull only proposes; Tim confirms. The earlier Claude scheduled task stalled on a permission prompt on its first night (Oct 14); its stuck run was stopped and the task is paused. A backup check in Anthony's main Claude session at 1:31 AM runs the pull only if no Pull State row was written that night, and it covered Oct 14 to Oct 16. The new runner was proven today: a request filed at 3:44 PM ran at 3:47 PM (154 logs read, 0 errors, Pull State row 11 written). The scheduler that files it automatically was installed at 3:49 PM, so tonight is its first unattended run.

How you use it

Nothing: it runs by itself. Results show in the Delay Log (Pull State, Delay Log and Candidates tabs) and in the bridge request Tim's Claude can read.

Who has access

Tim (owner of the spec); the Studio files and runs it

Watch out: Today's run found that JobTread has no daily logs dated after Oct 8, and nobody has filled in the delay fields on any log. So the pull currently has nothing to propose. The 1:31 AM backup lives in one Claude session and lapses when that session closes or its 7-day limit runs out (around Oct 18). The Studio watchdog now checks the bridge worker's run log for each night, from Oct 17, instead of the old paused task.

What it is

Every night it saves every open schedule item in JobTread for the active houses and flags any planned date that moved since the night before.

Why we built it

Part of Tim's weekend runbook. It gives the delay system a dated history of the schedule, so slips can be proven and the forecast has fresh data.

How it works

A worker reads JobTread through the Studio's JobTread connection. It appends about 492 rows for 22 houses to the 'Schedule Snapshots' tab of the Delay Log, writes any moved dates to the 'Slips' tab, then re-reads and checks what it wrote. It has run every night since Oct 10. The last run was 1:26 AM today, with 0 slips; schedules have not been edited for several days.

How you use it

Nothing: it runs by itself. Look at the Schedule Snapshots and Slips tabs in the Delay Log.

Who has access

Tim and anyone who can open the Delay Log

Watch out: This is NOT a standalone scheduled job. It is a timer inside one open Claude session on the Studio. Those timers stop if the session closes and expire 7 days after they were made. This one was made early on Oct 10, so it is due to lapse about now. The Monday forecast waits for this snapshot.

What it is

A read-only forecast of each active house's Final inspection date and its Estimated Completion date (Final plus 10 days). It runs every Monday and keeps a weekly history.

Why we built it

Tim decided on Oct 16: 'switch on the forecast'. He wants a simple weekly sheet of forecast completion dates and never wants dates applied automatically.

How it works

Every Monday at 5:50 AM it waits up to an hour for that night's JobTread snapshot. It places each house on the inspection chain (past UES inspection passes plus the FPL power step). It then adds the typical time for each remaining step, using the last 5 finished houses where possible. Results go into the 'Forecast - Estimated Final (Weekly)' sheet in Delay System Control, a history row is added, and a Gmail draft goes to Tim's inbox labelled To Do. It never writes to JobTread or the GB Inventory Sheet. It builds on a first read-only study (Oct 13, the 'Forecast Intervals - First Pass' sheet). That study found that JobTread 'Done' dates before Oct 9 were set in one bulk edit, so only the UES inspection chain is reliable enough to forecast from.

How you use it

Tim reads the Monday draft in his inbox and the weekly sheet in Delay System Control. Anyone can open the sheet.

Who has access

Tim; anyone with access to Delay System Control

Watch out: So far the only output is a draft made by hand on Oct 16; the first automatic run is Monday Oct 19 at 5:50 AM. 19 of 22 houses are low confidence. Six houses have no known FPL power branch. The forecast depends on the nightly snapshot, which may lapse (see that entry). The builder flagged a few data gaps for Tim, including which Weslaco lot FPL means by '1036 Weslaco'.

What it is

Weekend-run tools that proposed a new Est. Completion date for the 10 in-progress houses and would copy any date Tim approved into the GB Inventory Sheet.

Why we built it

Tim's original weekend runbook asked for forecast proposals that he approves with a Y at the Monday meeting.

How it works

The proposals are frozen in the Delay Log's Forecast tab, one row per house. A timer in Anthony's main Claude session checks every 15 minutes on weekdays, 7 AM to 9 PM, for a Y in the approval column. Only then would it write that one date into the Inventory sheet, re-reading the cell before and after. Every check so far shows all 10 houses 'not approved', so nothing has been written. The last check was 5:28 PM today.

How you use it

Ask Anthony. Tim would mark Y in the Forecast tab, but see the caveat.

Who has access

Tim (approves), Anthony

Watch out: On Oct 12 Tim changed his mind: no Y approvals and no automatic writes to the Inventory sheet. Forecasts are just handed to him weekly. The weekly forecast above now does that. The watch timer is still running, though, and its note cites an Anthony approval for the Inventory write on Oct 12. Leftover tool; should probably be retired.

What it is

Over one weekend the Studio worked through Tim's job queue for the delay system (J34 to J52) from his Weekend Runbook. It also prepared his Monday package.

Why we built it

Tim handed the weekend build to Anthony's Claude ('no approvals needed to build; build and present'). Anthony chose to run it all from the Studio.

How it works

A master Claude session on the Studio launched worker runs for each job and checked their results. Main jobs: a BuilderTrend-vs-JobTread checkbox reconcile across 27 houses (291 differences; 83 fixed, 208 held for Tim), JobTread link fixes, Registry v2, the nightly snapshot, the Final Review page, report cards, PM brief drafts, Forecast v2 and a full dry run. JobTread changes went in as Rich Pierce, because Tim's own JobTread grant can't be reached from the Studio. Google writes went through the Studio's relay. The Monday package and two PM briefs were saved as Gmail drafts from anthony@ to Tim; Anthony sent the package on Oct 12. Tim's open decisions are listed in one report. A copy of the weekend's output is indexed in the knowledge base.

How you use it

Read-only now. The results live in the Delay Log and Delay System Control; ask the Studio's Claude for detail.

Who has access

Tim, Anthony

Watch out: The PM morning briefs (J41) and the Monday package were one-off drafts. Nothing makes them every day or week. Tim's decision list from that weekend is still mostly open. Its items include partial passes, cancel chains, 5 'bypass' links, the Trainfo Plumb date and the daily carrying cost. The forecast-method conflict is in the open questions.

What it is

A private link that lets Tim's own Claude on his MacBook talk to the Studio's Claude. It can ask quick read-only questions or file whole jobs for the Studio to build.

Why we built it

Anthony, Oct 12: when Tim's Claude asks for a build, it should go to the Studio's Claude so both can talk. It also saves Tim's own Claude usage, because the Studio's account does the heavy reading and building.

How it works

Tim's Claude runs a small command, gb-studio, that connects to the Studio over Tailscale (a private network). The login key only works from Tim's Mac and can only run the bridge program, which accepts a fixed set of commands and logs every call. 'ask' gets an answer from a locked-down read-only Studio Claude, usually in 10 seconds to 3 minutes. That Claude can read only the ops code, the delay index, Tim's workspace and the requests folder, with no web, email, Drive or JobTread. 'request' files a job; 'followup' adds to one; 'status', 'list', 'log' and 'watch' show progress. Usage so far: 17 jobs filed, 22 follow-ups and about 3,400 status checks; the ask lane has been used only 6 times. Limits: 200 new jobs and 500 follow-ups a day, 30 asks an hour.

How you use it

Tim (or his Claude) types requests in his own Claude; it files them with gb-studio. Anthony can see them with the bridge admin tool on the Studio.

Who has access

Tim's Claude on Tim's MacBook only; Anthony administers on the Studio

Watch out: It only works while the Studio is on and logged in. The key is pinned to Tim's Mac's network address, so the bridge stops if that address changes (for example if his Mac is re-added to Tailscale). Text sent by Tim's Claude is treated as data, never as instructions to bypass the limits. Turning it off means removing one line from the Studio's login keys. The earlier 'text Anthony to approve' app was switched off Oct 14 and nothing texts anyone now.

What it is

Each job Tim files gets its own Claude on the Studio that does the work end to end and posts progress back to Tim.

Why we built it

Anthony, Oct 13: 'link his tasks to individual terminal Claude chats ... live updates'. Then Oct 14: 'make it automatic and seamless' and 'I don't need to OK everything he does'.

How it works

A dispatcher checks for new requests every 20 seconds and starts one worker Claude per request, up to 3 at once, each in its own session. Workers run in automatic mode: Tim's request is the approval. They use the JobTread connection (as Rich Pierce), Google Drive and Sheets through the Studio's relay, Tim's mailbox tool and BuilderTrend read access. A safety hook checks every action and refuses five things outright: payments, typing or reading passwords and keys, permanent deletes, pushing or releasing Anthony's systems, and editing the bridge itself. A refused step is skipped and listed in the reply; nobody is asked or texted. Progress posts to Tim use a fixed word list so no free text leaks. Workers back up anything before overwriting it. A worker closes 20 minutes after its job is done; a follow-up wakes it again.

How you use it

Tim files a request from his Claude; the worker replies on that request. Anthony can list or open live workers with the bridge admin tool.

Who has access

Requests come only from Tim's Claude (and the Studio's own nightly filer)

Watch out: Workers run on the Studio's Claude login. Workers sometimes started 'Not logged in' (seen 4 times on Oct 14); the dispatcher restarts them up to 3 times. The hook has had false alarms that were later fixed, so a refused step can mean a hook bug rather than a real limit. Releases of the shared ops app stay with Anthony.

What it is

Since Oct 16 Tim has owner authority over every Gibraltar production system and tool, built now or later. His requests never wait for Anthony.

Why we built it

Anthony on Oct 13: 'NONE OF THIS NEEDS MY PERMISSION IF HE SAYS IT LET IT JUST GO'. Extended Oct 16 to owner level on everything.

How it works

Tim's knowledge-connector key was raised to owner (nothing hidden, every inventory column, the closings tool). His ops dashboard login is now Administrator (checked: role 1). His mailbox rights were widened (see the mailbox entry). The Studio's Google relay can now edit anything in the Gibraltar shared drive plus the GB Inventory Sheet, but never delete, trash or change sharing. The worker rules and Tim's Studio workspace were updated to 'no OKs'. An update script for Tim's Mac was placed in the shared drive (Tim Emails > Claude setup).

How you use it

Tim simply asks, in his Claude or at ops.gibraltarbuilders.info (Administrator). Nothing to do for Daniel.

Who has access

Tim

Watch out: Claude's own limits still apply to Tim: no payments, no passwords, no permanent deletes, and no releases of the shared ops app. The bridge log shows Tim's Claude never fetched the Oct 16 notice, so the update on his Mac (new gb-studio version and the 'no approvals' text in his Claude settings) is unconfirmed. Tim has no shell or screen access to the Studio itself.

What it is

A Studio tool that lets Tim's jobs work in his Gmail (tim@gibraltarbuilders.com). It can also read the other company inboxes.

Why we built it

Tim's Oct 13 email asked for his Studio jobs to use his mail (vendor prices, credit applications, follow-ups). Anthony said 'check tims email and fix that please'.

How it works

The tool sends each request to a fixed script on the ops server, which uses the company's Google service account. For tim@ it can search, read, download attachments, add or remove labels, create drafts and, since Oct 16, send. It refuses trash and spam labels and never deletes. Any other @gibraltarbuilders.com inbox can be searched and read but never sent from. That includes Daniel's, scheduling@ and billing@.

How you use it

Used by Tim's Studio workers and routines; Tim asks his Claude. Nobody else uses it directly.

Who has access

Tim (through his Studio requests); Studio sessions

Watch out: Daniel should know that Tim's Studio jobs can read every company inbox, his own included. The live ops app was not restarted after the Oct 16 change, so the ops screens still treat tim@ as read-only until the next release. The tool itself already works.

What it is

Every Monday it sorts Tim's sent mail into Outstanding (someone owes Tim) and To Do (Tim owes someone). For threads waiting a week or more it drafts a short 'Any updates? Thanks!' reply for him to send.

Why we built it

Tim asked (Oct 13) to move this weekly routine from his MacBook to the Studio so it runs reliably every Monday.

How it works

At 6:30 AM Monday it reads Tim's sent mail since the last run. It sorts new threads with fixed rules (when unsure: Outstanding) and applies the Gmail labels Outstanding and To Do. It creates one threaded draft per open Outstanding thread that is 7 or more days old, and takes the label off threads that were answered. It never sends or archives. A report goes to the 'Tim Emails / Email Routine' folder in the shared drive and is posted to Tim's bridge request. The Oct 14 dry run found 15 drafts it would make, 19 new Outstanding and 8 new To Do, mostly replies to Tim's 'Do you accept credit cards?' supplier emails.

How you use it

Tim opens Gmail on Monday and sends or deletes the drafts. Labels Outstanding and To Do show what is open.

Who has access

Tim

Watch out: Only a dry run so far (Oct 14); the first real run is Mon Oct 19 at 6:30 AM, and the census still lists 'check the first run'. Sorting uses fixed rules, not Claude: a Claude sorter was refused by the safety limits and was set aside. If Tim's MacBook still runs its own Monday task, the two could make duplicate drafts. The builder asked Tim to switch the Mac task off before Oct 19. The Studio watchdog checks that Monday's run produced its report.

What it is

A report-only Monday check of permit and inspection status for the 9 homes that do not yet have a building permit, read from Palm Bay's permit portal (IMS) and the private inspection provider (UES).

Why we built it

Part of Tim's overnight batch of Oct 12 (Job D): he wants to know each week which permits moved.

How it works

A Claude scheduled task on the Studio runs a script that reads the IMS and UES robot browser windows already open on the Studio. It writes a table per home (building permit, public works permit, inspections) and compares with last week. The report is posted to Tim's bridge request as a reply plus a one-line progress note. It never writes to JobTread, sheets or the portals, never logs in and never accepts the UES disclaimer.

How you use it

Tim's Claude reads the reply on his bridge request (gb-studio status or watch).

Who has access

Tim

Watch out: It has never run for real: there is only the Oct 13 test report, and the task shows no past runs. It runs as a Claude desktop scheduled task, so the Claude app must be open on the Studio, and it is the same mechanism that stalled on a permission prompt for the nightly pull. It also needs the robot windows signed in. In the Oct 13 test IMS was read but UES was not: the UES window shows a disclaimer only Tim may accept, and it comes back on every reload. The IMS window has looked signed out since Oct 13 evening (still so at 6:11 PM today), so Monday's report will probably say IMS was not read unless someone signs in by hand first. Two homes in other jurisdictions are skipped.

What it is

Two Chrome windows kept running on the Studio: one for UES (private building inspections) and one for Palm Bay IMS (city permits). Studio robots can read inspection data from them.

Why we built it

The inspection portals have no API and their sign-ins expire. Keeping a signed-in window open lets the delay system and permit check read inspection dates without anyone typing a password.

How it works

A person signs in by hand once in each window. A background script keeps each session alive with a quiet page refresh every 10 minutes; it never types or clicks on the portal. Robots attach to these windows to read only. They have run continuously since Oct 12, 3:21 PM.

How you use it

Ask Anthony. Someone at the Studio must sign in by hand when a window is signed out, and Tim must accept the UES disclaimer himself.

Who has access

Studio robots; the UES login is Tim's

Watch out: UES: the session is alive (keep-alive answered at 6:11 PM today), but the disclaimer blocks reading until Tim accepts it. IMS: signed out since Oct 13 evening (still so this evening). Today these windows are read only by the Monday permit check (first real run Mon Oct 19) and by jobs run by hand. The robots that would feed inspection dates into the Delay Log automatically are built but not installed (the census lists the same).

What it is

Every Friday it reads suppliers' answers to Tim's Oct 13 email 'Do you accept credit cards?' and fills one row per supplier in a log sheet. The row records which cards they take, fees, limits and discounts.

Why we built it

Tim's request (Oct 13): collect the replies to his 20 supplier emails each Friday until his card-rewards spreadsheet is built.

How it works

At 3 PM Friday it searches Tim's mailbox (read only) for replies on the 20 threads and card-related mail from the same suppliers. It fills the row for anyone who newly answered and adds later replies to notes. It backs up the Log tab first. Rows it fills itself are marked for Tim to check. From Oct 23 it marks suppliers with no answer 'No reply yet'; it never chases anyone. Today's 3 PM run found 1 new reply: 18 answered, 3 still silent.

How you use it

Open the supplier card log sheet in the shared drive's Finance/Pricing folder. To stop it, set the Run tab's switch (cell B1) to OFF.

Who has access

Tim; anyone with access to Finance/Pricing

Watch out: It never copies card or account numbers. Runs only while the Studio is on and logged in.

What it is

A series of Studio jobs that filled in Tim's 'Gibraltar Vendor Directory (Master)' sheet and a per-house vendor table.

Why we built it

Tim: 'For any vendor categories ... that don't already have 3 options minimum, add in the backup trades ... color code the ones you add so I know to review.' Follow-ups asked for prices and credit status.

How it works

Workers used verified sources only. They added backup trades to bring each category to 3, shaded for Tim's review; Tim approved all but one, which was replaced with a tile installer. They wrote 36 high-confidence contact cells and skipped a dissolved company. They filled the Pricing column and its source column from quotes in the shared drive and Tim's email, and set a 'Credit Active' status on supplier backups. They also added an 'Expanded by House' tab to the Backup Vendors Tracking sheet (26 houses by about 60 trades, deviations highlighted). Every write was backed up and re-read.

How you use it

Open the Vendor Directory in the shared drive under Finance/Pricing/Data. New work goes through Tim's Claude.

Who has access

Tim; anyone with access to Finance/Pricing

Watch out: Medium-confidence contact names (mostly from state company records) were not written. Shaded cells are additions Tim still owns reviewing.

What it is

Two finance jobs for Tim. One collected every credit application document from his mailbox into a renamed, indexed 'Credit Applications' folder. The other built a 'Card Rewards Projection' sheet estimating yearly card points if suppliers are paid by card.

Why we built it

Tim: 'scrape through all of my tim@ emails for credit application documents and save them all to the shared drive in the pricing folder'. Then (Oct 16): estimate the points per year assuming 50 houses, split like the Gibraltar Savings document.

How it works

Credit applications: 18 files uploaded unchanged, renamed for browsing, plus an index sheet showing each supplier's status (approved, pending, not submitted, no paperwork). Payment and wire forms were left out on purpose. Card rewards: the sheet copies the layout of 'Gibraltar Savings (Aligned)'. It puts each trade on a supplier that takes cards with no fee where one exists, and lists ties and judgment calls for Tim.

How you use it

Both are in the shared drive under Finance/Pricing. Tim edits the Scenarios cells to change supplier choices.

Who has access

Anyone with access to Finance/Pricing

Watch out: The credit forms contain sensitive company details, and the job asked Tim who should have access to Finance/Pricing. An earlier empty 'Credit Applications' folder and an old index copy sit in Delay System Control (marked superseded; never deleted). The projection puts about $850K a year of spend on Builders FirstSource, whose no-fee card terms are 'cash customer only'. That should be confirmed.

What it is

The Studio's route for creating and editing Google Sheets, Docs and Drive files. Tim's jobs use it.

Why we built it

Built for the weekend delay run so the Studio could write Tim's sheets; later widened so Tim's jobs can work anywhere in the shared drive.

How it works

Studio tools pass each request to a fixed program on the ops server. That program acts as rich@ through the company's Google service account. Guards: no deletes, trash, moves out or sharing changes. No Delay Log writes between 11:40 PM and 12:40 AM. Since Oct 16 it may edit any file in the Gibraltar shared drive plus the GB Inventory Sheet. Every call is logged on the server. A companion tool can create Gmail drafts in anthony@ addressed to Tim (used in the weekend run); it never sends.

How you use it

Nothing directly. Tim's requests and Studio sessions use it.

Who has access

Studio Claude sessions and Tim's workers

Watch out: Edits show up in Drive as made by rich@, not by Tim or Claude.

What it is

A Google Sheet with one row per closing, built from the title companies' settlement statements. It also fills empty cells in the GB Inventory Sheet.

Why we built it

Anthony, Oct 13: 'as new closings happen, we need a record Daniel can easily access'.

How it works

Several times a day it searches kelsey@, billing@ and anthony@ for settlement-statement PDFs and reads them; scans are run through text recognition. Claude on the Studio pulls out the closing data, with no internet or tools during that step. It checks every dollar amount word for word against the PDF and merges versions into one row per title file. Tabs: Home Sales, All Closings, Needs Review, Missing Statements, Dashboard Check, Inventory Sheet Fills, About. Statement copies go to the Settlement Statements subfolder. Today it holds 186 closings, 82 of them home sales from 2022 to 2026. It also fills GB Inventory Sheet cells that are empty and formula-free, from signed statements only, in the sheet's own format, and never overwrites. Today it wrote 3 cells, and 2 are held because the sheet disagrees with itself.

How you use it

Open 'Gibraltar Closings Log' in the shared drive under Office Opps/Closings Log. Or ask your Claude through the Studio connector; the 'closings' tool works for owner keys, including Daniel's and Tim's.

Who has access

Daniel, Tim and anyone with access to Office Opps; owner and dev keys on the Studio connector

Watch out: It failed every run from Oct 15 6:45 AM to Oct 16 1:05 PM: the Claude reading step timed out while the Studio was overloaded. It has run cleanly since 3:12 PM today. 3282 Westmore and 2166 Santa Maria lot dates are held for a person because the sources disagree. Lot purchases on the inventory sheet look like 'Close' statuses but are not home sales.

What it is

Once a day it emails Daniel the homes listed on the sales site that Gibraltar's own records contradict: sold, held internally, closed in FD Funding, or a closing in progress while listed Available.

Why we built it

Daniel, Oct 13: 'build the daily dashboard check and email me the flags'.

How it works

After each Closings Log sync it reads the sales site's home and lot list, the Closings Log, FD Funding statuses and recent kelsey@/anthony@ mail whose subject names the home. It flags contradictions and marks new ones against yesterday. It publishes a 'Dashboard Check' tab and emails daniel@ from anthony@ once a day. It never edits the sales site or the inventory sheet. Today's email went out at 3:14 PM with 2 flags, 0 new.

How you use it

Read the daily email, or the Dashboard Check tab in the Closings Log.

Who has access

Daniel (recipient)

Watch out: No email went out on Oct 15 because the closings sync was broken (failing from 6:45 AM Oct 15 until it was fixed at 3:12 PM Oct 16). Today's came at 3:14 PM instead of 6:45 AM. The census item 'daily email to Daniel stopped Oct 15' is now out of date but still shows as open on the Progress tab. The sales site rebuilds from the GB Inventory Sheet every 4 hours, so a stale site usually means a stale sheet. 2720 Ohio is probably an internal hold but is not on the site.

What it is

A robot that copies the superintendents' daily-log photos out of BuilderTrend. It attaches them to each home in the ops Homes & Stages page with a stage, and files them in the home's Stage N folder in Drive.

Why we built it

Asked Oct 5: walk each home's stages in BuilderTrend, save the photos, and attach them to ops and the right Drive folders at the same time.

How it works

A Chrome window signed into BuilderTrend (as rich@) reads BuilderTrend's own data feed; it never clicks around. New photos get the home's current stage from the log date. Undated or unclear ones are guessed by a free local image model and marked 'Needs a stage' for a person to approve or move; it learns from those choices. The ops server then files each original photo in Sales/Projects - Sales/<address> - Sales/Stage Update Photos/Stage N, creating folders when missing and never deleting. 1,479 photos across 30 jobs are known.

How you use it

At ops.gibraltarbuilders.info, open Homes, then a home's photos: approve, move ('Wrong stage?') or remove ('Not a build photo'). Drive folders fill themselves.

Who has access

Ops users with the Homes tab (Tim, Kody, Jacob, Daniel, admins)

Watch out: The 6 AM run failed Oct 14, 15 and 16 (Chrome did not start while the Studio was overloaded). A hand run at 3:11 PM today worked and picked up 14 new photos, 5 needing a stage. Tomorrow's 6 AM run is the first unattended check since. BuilderTrend records no stage tags, so stages are inferred. The older Drive-folder photo intake is a separate path and still fails every 15 minutes: its Google login on the Studio expired.

What it is

A folder on the Studio where a Claude session behaves like Tim's own Claude, with his rules and memory, for when Tim sits at the Studio.

Why we built it

Tim at the Studio, Oct 9: 'I need you to be just like that Claude.'

How it works

Its instruction file holds Tim's style rules, his owner-level access and the delay-system rules. It pulls in read-only copies of his Mac's Claude memory and workspace instructions. Those copies are refreshed every 15 minutes from the Mac Mini, which collects them from Tim's Mac; last refresh 5:35 PM today. A matching system prompt for Tim's claude.ai preferences is kept alongside.

How you use it

Tim opens Claude Code in that folder at the Studio.

Who has access

Whoever sits at the Studio and opens Claude there for Tim

Watch out: The security review of the bridge (Oct 12) flagged one thing that is still open. Sessions in this folder load Tim's synced notes as instructions, so anything written on Tim's Mac becomes an instruction here. Anthony left it as designed.

What it is

Tim's Claude memory, chat transcripts, Cowork sessions and his Desktop/CLAUDE workspace are copied, read-only, to the Mac Mini's knowledge service. Anthony's Claude can search them there.

Why we built it

Anthony, Oct 9: 'make sure tims computer is sharing its claude knowledge base with the mac studio ... we just need all access ... index everything'.

How it works

Tim ran a setup script on his Mac that allows only the Mini to copy those folders, read-only. The Mini pulls about every 15 minutes (46 files changed in the last day). The first full pull was about 1.1 GB. The knowledge service's 'tim' source now holds 107 documents, searchable by the Studio's Claude. A summarizer that turns his raw chats into short session records has been built but not scheduled (17 record files exist, only 2 of them session records).

How you use it

Ask the Studio's Claude ('what did Tim's Claude decide about X'). Daniel's connector does not include this source.

Who has access

Anthony's Studio and Mini Claudes

Watch out: His Mac must be awake and online. Git history is excluded on purpose. The session-record summarizer is waiting on a decision about how strictly it filters out personal content: in tests it let some personal-finance material through.

What it is

Tim's own Claude can search Gibraltar's Drive knowledge base, the live GB Inventory Sheet, the systems guide and the Closings Log through the Studio's knowledge connector.

Why we built it

Anthony, Oct 9: 'set him up with the knowledge connector and make sure all his Claude chats going forward use it.'

How it works

A personal key for Tim's Claude was created on the Studio and installed on his Mac by the Oct 9 setup file. Calls go over the private Tailscale link to the Studio's read-only connector. Since Oct 16 his key is at owner level (nothing hidden). 30 tool calls so far, mostly document searches.

How you use it

Tim asks his own Claude; it uses the 'gibraltar' connector automatically.

Who has access

Tim

Watch out: Read-only. Anthony can revoke the key at any time.

What it is

Tim's MacBook is linked to the Studio and the Mac Mini over Tailscale, a private encrypted network. The bridge, the knowledge connector and the notes sync all depend on that link.

Why we built it

Needed so Tim's Claude can reach the Studio, and the Mini can copy his notes, without exposing anything to the internet.

How it works

There are three separate Tailscale networks: the Studio's, Anthony's (with the Mini) and Tim's personal one. They are joined by sharing individual machines. On Oct 9 the Studio was shared with Tim and Tim's Mac with the Studio, both ends accepted. Over the office network the link is direct and fast.

How you use it

Nothing: it runs by itself. If it breaks, ask Anthony; undoing it means removing the machine share in each Tailscale admin console.

Who has access

Tim's Mac; Anthony manages the Studio's and Mini's sides

Watch out: Re-adding Tim's Mac to Tailscale changes its address and would break the bridge key until it is updated.

What it is

Every night the Studio copies all of Gibraltar's important data, keeps 30 days of copies, and puts a second copy on the office Mac Mini. At 8 AM a separate check confirms that last night's backup actually ran.

Why we built it

The FD Funding ledger and the user accounts exist only on the ops server. Before this was hardened, a failed backup looked just like a good one: on two nights in August the wrong database was backed up and nobody noticed.

How it works

At 2:30 AM it takes a consistent snapshot of the production ops database (FD Funding advances and properties, users, leads) and checks the row counts against the live system. It also exports the CRM database from Cloudflare, downloads the FD Funding Control and GB Inventory spreadsheets as Excel files, and bundles the full code history and the production settings file. New closing documents and stage photos are added to running copies that are never pruned. The night's folder is then copied to the Mac Mini. If any piece fails, Anthony gets an email. The 8 AM check emails him if the backup did not run at all, which covers cases like a sleeping Mac or a lost schedule.

How you use it

Nothing: it runs by itself. To restore anything, ask Anthony.

Who has access

Anthony, through the Studio and Mini logins. Alerts go to Anthony only.

Watch out: Healthy: Oct 16's run finished at 3:11 AM with 0 failures (475 advances, 184 properties, 9 users, 697 leads verified), and the 8 AM check has reported OK every day since Oct 2. There is no off-site copy: the Mini sits in the same building, so a fire or theft would take both. The Mini's disk is not encrypted (FileVault is off), and its copy includes sensitive settings and customer data. A restore-drill script exists but is not scheduled.

What it is

Every morning the Studio refreshes Gibraltar's list of about 8,500 vacant lots in Brevard and sends it to the ops Lots (Lot Finder) and Metrics tabs. It also reads new building-permit cards from the Drive and adds them to ops.

Why we built it

To keep the Lot Finder call list and the Metrics tab current from free public data. The lots revamp target is about 9 lots bought per month.

How it works

It scans Brevard County's public parcel map for vacant lots, adds new ones and refreshes the ones we track (owner changes, a house built, a lot gone). Then it checks Palm Bay water and sewer availability, build-out density, and flood, wetland and scrub-jay habitat against county map layers. It also reads tax-deed pages that a person has saved into an inbox folder, and clears owner phone numbers from lots that were built or sold more than 30 days ago. Finally it pushes all of this, plus the permit cards from the Drive copy, to the ops server and records a health summary. It uses only free sources and never contacts anyone.

How you use it

ops.gibraltarbuilders.info, Lots tab (call list) and Metrics tab. Nothing to run.

Who has access

Results are visible to anyone with the Lots or Metrics tab in ops. The job itself is Anthony's.

Watch out: Healthy: Oct 16's run took about 27 minutes, every step passed, 8,532 lots and 61 permits were pushed. No new county lots have appeared in 3 days because the county list is fully caught up; new lots now come from tax deeds and listings. The wetland and lot-shape checks are waiting on Anthony's first full hazard pass, so nearly every lot still shows wetland as 'unchecked'. The tax-deed step only works when a person saves realTDM pages into its inbox, since realTDM blocks scripts.

What it is

Two older Studio jobs whose schedule lines are commented out. The Friday paid sweep was switched off on purpose. The invoice sender, still named the 'Dext send agent', was not a leftover: it was the Studio half that delivered vendor invoices from billing@ to QuickBooks Online's receipt capture, and it has stopped.

Why we built it

The Friday sweep spends money on paid data services without a per-batch OK. The invoice sender's line was marked 'paused Aug 14, invoice capture moved off Dext', but invoices only changed destination (from Dext to QuickBooks); they still needed delivering.

How it works

The Friday 8 AM 'accountability sweep' re-checked six realtors' 2026 deals against our leads using paid property data (Datafiniti and REAPI) and updated the dashboard. Anthony paused it on Oct 9 until a spending confirmation step ships. On Oct 16 Daniel's seat committed a free replacement for its 'did this lead buy a home' half, using free county data (dry run only; the schedule still points at the paused paid job). The invoice sender emptied the ops server's invoice queue every 5 minutes and forwarded each invoice by email. Its log shows 98 invoices delivered to QuickBooks between Aug 14 and Aug 26, and nothing after Aug 26 at 2:05 PM. 126 invoices queued between Aug 26 and Sep 15 have never been sent, and the billing@ reader on the ops server has failed every run since Sep 15.

How you use it

Sweep: ask Anthony, or Daniel for the free replacement. Invoices: until the pipeline is repaired, invoices sent to billing@ have to be forwarded to QuickBooks by hand. See the 'Vendor-invoice pipeline' entry.

Who has access

Anthony

Watch out: Even if the sweep were switched back on, both paid sources would fail today: REAPI answers 'payment required' and the office router's security filter blocks Datafiniti. A top-up or a router change would make it start spending again. The invoice sender's stop was not flagged anywhere: the Studio watchdog does not watch it, and the Systems Guide still says invoice forwarding is live.

What it is

Prepares every construction photo that enters ops Homes & Stages so it can be shown: it converts iPhone photos, removes location data, makes web-sized copies and keeps the original.

Why we built it

iPhone photos use a format (HEIC) that the ops server cannot decode, while the Mac can do it for free. Without this step, uploads sit at 'processing...' forever.

How it works

Every 30 seconds it asks the ops server for photos waiting to be processed. It decodes each one with the Mac's built-in converter and strips the GPS and other hidden data so a buyer's address is never shipped inside a picture. It makes display copies, sends them back to ops, and archives the original on the Studio, where the nightly backup includes it.

How you use it

Nothing: photos uploaded in Homes & Stages, or brought in by the BuilderTrend robot, simply turn from 'processing' to ready.

Who has access

Anthony. The photos are visible to whoever has Homes & Stages.

Watch out: Healthy: it processed the 14 new BuilderTrend photos at about 3:12 PM on Oct 16 with no errors.

What it is

Every 15 minutes this job is supposed to pick up photos that someone drops by hand into a home's Drive 'Stage Update Photos / Stage N' folder and bring them into ops Homes & Stages.

Why we built it

So stage photos can be added just by dropping them into Drive, without logging into ops.

How it works

It gets a short-lived Google pass through the Studio's Google Cloud sign-in, reads the Stage folders (read-only), and sends new photos to ops, where they wait unapproved until a person approves them.

How you use it

Would be: drop photos into the home's Stage folder in Drive. Today nothing comes through.

Who has access

Anthony

Watch out: Every run since Sep 18 has failed ('could not mint a Drive token': the Google sign-in behind it expired), which is about 4 weeks. The quick fix is one sign-in command that Anthony runs at the Studio. Because that sign-in expires again within days, the lasting fix (already on the open-items list) is to move it to the newer Google connection. The BuilderTrend robot now carries the main photo flow, so this only matters for photos added to Drive by hand. The new Studio watchdog does not watch scheduled jobs like this one, so it is not flagged anywhere.

What it is

The piece that would email buyers their home's stage updates once a person has drafted and approved one in ops. It runs every 5 minutes but is in test mode, and the program is on hold.

Why we built it

The ops server's host blocks outgoing email, so the Studio was set up to deliver the approved update emails.

How it works

Every 5 minutes it checks the ops server for approved stage updates and sends them from scheduling@. A test setting currently redirects every recipient to Anthony's personal inbox, so no buyer receives anything. It is configured to send through Resend, whose sending domain has never been verified, and the last real attempt (1387 Platt Ave, Sep 15) failed for that reason.

How you use it

In ops, Stage Updates: someone drafts and approves an update. Today nothing reaches buyers.

Who has access

Anthony

Watch out: On hold since Sep 22 at Rich's word ('until we finalize it with Tim'). The outbox is empty. A replacement that sends through Google from the ops server is built but switched off. The Platt Ave update still needs a person to re-draft it.

What it is

It reads every settlement statement that title companies email to Gibraltar and keeps the 'Gibraltar Closings Log' Google Sheet up to date with one row per closing. It also emails Daniel a daily list of sales-site listings that the closing records contradict.

Why we built it

Anthony, Oct 13: 'as new closings happen, we need a record Daniel can easily access.' Built for Daniel's closings reports.

How it works

It searches the kelsey@, billing@ and anthony@ mailboxes (read-only) for settlement-statement PDFs from the last 5 days and turns them into text, reading scans with text recognition. Claude (Anthropic's Sonnet model) reads each statement once. Every dollar figure is then checked word-for-word against the PDF, and unclear ones are flagged for review. It rebuilds the sheet tabs (Home Sales, All Closings, Needs Review, Missing Statements, Dashboard Check) and files statement copies in Office Opps / Closings Log. It also fills empty cells in the GB Inventory Sheet, only when the cell is empty and the statement is signed, and never overwrites anything.

How you use it

Daniel opens the Gibraltar Closings Log sheet (shared drive, Office Opps > Closings Log), reads the daily 'dashboard check' email, or asks his own Claude, which is connected to the Studio and has a 'closings' tool.

Who has access

Daniel and anyone the sheet is shared with; Daniel's and Anthony's Claude through the connector.

Watch out: Every run failed from 6:45 AM Oct 15 to 1:05 PM Oct 16 (7 runs) because the Claude reading step timed out while the Studio was overloaded, so no daily email went out on Oct 15. It was fixed at 3:12 PM on Oct 16 and today's email went out at about 3:14 PM (2 flags). The census items 'fix the Closings Log sync' and 'daily email to Daniel stopped Oct 15' are now out of date but still show as open on the Progress tab. To date: 627 PDFs, 186 closings, 82 of them home sales. Statement contents are sent to Anthropic to be read.

What it is

Keeps a private, searchable copy of the whole Gibraltar shared Drive on the Studio. This is what Claude, the Discord bot and the Studio connector search when someone asks 'what do we have on X'.

Why we built it

Rich asked for the AI to 'know' the Drive without paying for tokens. Anthony chose local search, which can cite the real file, over training a model, which freezes and invents details.

How it works

Every night at 2:30 AM it copies the Drive (read-only Google access, about 10,600 files and 14 GB), pulls the text out, reads scanned PDFs in the important folders, adds the Studio's own Claude conversation transcripts (searchable on the Studio only), and indexes everything with a local AI model (315,021 passages from 5,651 documents). It then publishes tiered snapshots: what staff may see, what only leadership may see, and what is never shown. It also redraws the knowledge-base 'brain' map. Between nights, a live refresh checks Drive every few minutes and folds changes in, but it waits while the nightly run is going. A small web server shows the brain map on the Studio only.

How you use it

Ask Claude through the Studio connector (Daniel's Claude has it), mention the Hermes bot in Discord, or ask Anthony. On the Studio itself, Claude uses the 'kb' search command.

Who has access

Anthony on the Studio; staff and leadership through the bot and connector, by tier. Company login sheets are never indexed.

Watch out: The nightly run used to take about 20 minutes. Oct 14's ran about 29 hours (2:30 AM Oct 14 to 7:20 AM Oct 15), Oct 15's never started, and Oct 16's ran from 2:39 AM to 3:21 PM. Oct 11's Drive copy failed. While the nightly runs, the live refresh waits, so on Oct 16 no Drive changes reached search between 2:39 AM and about 3:23 PM, and Hermes, the Assistant and the connector answered from an early-morning copy. The slow runs fell during this week's overload, and load is back to normal; tonight's run will show whether it is back to its usual length. The snapshot publish was refused on Oct 14 and Oct 16 by a safety rule: the leadership-only tier grew more than 20% (1,058 to 1,362 documents). Until someone confirms that growth is expected, the bot and connector search the Oct 13 snapshot plus the live changes. About 1,325 scanned Production/Archive PDFs are still unreadable to search. The Drive copy uses a shared Google client that Google retires during 2026. Two files fail to read every night (one old .doc and one temporary Word file).

What it is

Free AI models that run on the Studio itself, so cheap or private jobs never send data out or cost money.

Why we built it

Local inference on this machine is free. Mechanical work (indexing, photo reading, summaries) goes here instead of paid cloud models.

How it works

The Ollama app serves four models on the Studio only: a small model that indexes the knowledge base, a vision model that reads BuilderTrend photos to suggest a build stage, and two general models used for free local answers ('ask') and for summarising Tim's Claude sessions into knowledge-base notes. Nothing it processes leaves the machine.

How you use it

Nothing: other jobs use it automatically.

Who has access

Only programs on the Studio

Watch out: Running and answering. It is started by the Ollama desktop app at login rather than as a background service, so after a restart someone must log in to the Studio before it, and the photo robot and knowledge-base jobs that depend on it, can work. Big jobs share the same chip, so a heavy run slows the others.

What it is

Every 15 minutes it reads Tim's bonus delay-tracking files (Job Board, Delay Log, candidates, his notes) from the Studio's Drive copy and keeps a current summary plus a running change log on the Studio.

Why we built it

Anthony, Oct 9: index Tim's delay-tracking system so the Studio always knows its current state without re-reading the sheets.

How it works

It reads only. It writes an INDEX file (current state) and a CHANGES file (newest first: job status moves, new or changed delay rows, nightly-pull runs, edits to Tim's notes).

How you use it

Nothing: Claude on the Studio reads it when asked about the delay system. Ask Anthony.

Who has access

Anthony and Claude on the Studio

Watch out: Healthy: last run 5:44 PM Oct 16; no changes since 4:14 PM.

What it is

A set of Tim's jobs that moved from his MacBook to the Studio: the nightly Delay Log pull from JobTread, a Monday forecast draft, a Monday 'who owes me a reply' email routine, a Friday supplier credit-card log, and a 15-minute copy of Tim's Claude notes.

Why we built it

Tim's rule: the Mac Studio always runs these, never his laptop.

How it works

The nightly pull is filed at 12:45 AM as a request to the Studio bridge, where a worker Claude runs it. The Monday 5:50 AM forecast waits for that night's JobTread snapshot, updates the read-only forecast and leaves a draft email in tim@. The Monday 6:30 AM email routine sorts Tim's sent mail, labels it and drafts 'Any updates?' follow-ups, and never sends. The Friday 3 PM job logs suppliers' replies about credit cards into a Finance sheet. The bridge itself runs Tim's requests as they arrive.

How you use it

Tim uses them through his own Claude and the Studio bridge.

Who has access

Tim (owner of these jobs) and Anthony

Watch out: The Friday supplier log ran at 3:03 PM on Oct 16 (21 rows, 1 new reply), and the notes copy last ran at 5:35 PM. The new 12:45 AM nightly pull runs for the first time tonight. The Monday forecast and the email routine run live for the first time on Mon Oct 19 (the email routine has only had a dry run, on Oct 14).

What it is

Two scheduled jobs that start a Claude session on their own. The Monday permit check reads the Palm Bay permit portal (IMS) and the UES inspection portal for 9 homes still without a building permit, then posts a report to Tim's bridge request. The nightly Delay Log pull task is switched off.

Why we built it

Tim's Job D (approved Oct 13): a weekly report-only permit check. The nightly pull task stopped working and was replaced.

How it works

Every Monday at about 6:15-6:22 AM, Claude runs a fixed check script against two robot Chrome windows that stay signed in to UES and IMS. A person signs those windows in by hand, and a keep-alive script touches each site every 10 minutes so the sign-in lasts. Claude then posts the report and a one-line progress note to Tim. It never logs in, never types a password, and never writes anywhere. The nightly pull task stalled on a permission prompt at its first run (Oct 14), which blocked every later run, so on Oct 16 it was paused and replaced by the 12:45 AM bridge job.

How you use it

Tim reads the Monday report on his bridge request. Nothing for Daniel.

Who has access

Tim and Anthony

Watch out: The Monday permit check has never run for real; there is only an Oct 13 test. Its first scheduled run is Mon Oct 19 at about 6:15 to 6:22 AM, and it needs the Claude desktop app open on the Studio. The IMS robot window has been signed out since 3:31 PM Oct 13 (still so this evening), so unless someone signs it in at the Studio, Monday's report will say IMS was not read. The UES window's session is alive, but a disclaimer that only Tim may accept blocks reading it. The two robot windows are not managed by the system scheduler, so they do not come back after a Studio restart. A Claude scheduled task can stall on a permission prompt the same way the nightly pull did.

What it is

Four repeating jobs that exist only inside one long-running Claude conversation on the Studio, left over from Tim's weekend delay run: a 12:47 AM JobTread schedule snapshot, a 1:31 AM backup copy of the nightly Delay Log pull, an estimated-completion watch every 15 minutes on weekdays from 7 AM to 9 PM, and a delay-system heartbeat every 6 hours.

Why we built it

Set up during the Oct 9-12 weekend run to keep Tim's delay system moving before permanent jobs existed.

How it works

Claude Code timers fire only while that conversation is open and idle. They stop if the session is closed, and they expire on their own 7 days after they were created.

How you use it

Nothing; ask Anthony.

Who has access

Anthony

Watch out: This is fragile. Tim's new Monday 5:50 AM forecast waits up to an hour for the 12:47 AM JobTread snapshot. If these timers expire (around Oct 17-19) or the session closes, that snapshot stops and nothing alerts anyone. Tonight the 1:31 AM backup pull and the new 12:45 AM nightly pull will both run, which should be harmless because the pull is designed to be safe to repeat.

What it is

Every 10 minutes it checks that the Studio's scheduled Gibraltar jobs and services are running and produced their results, and that nothing is eating the machine. When something new goes wrong, an AI reviewer writes a plain-English explanation.

Why we built it

Built Oct 16 after 24 stuck test helpers ran for 3 days (load about 95), while the BuilderTrend and Closings syncs failed and nobody knew.

How it works

It flags services that are down, timed jobs whose last run failed or never wrote its log, Monday and nightly jobs whose output file is missing, Tim's Delay Log not being pulled within 26 hours, runaway processes, overload, and low disk space. Since Oct 16 the nightly-pull check looks for the bridge worker's run log for each night (from Oct 17), instead of the old paused Claude task. The only thing it kills on its own is an orphaned Claude test process that has run hot for 30 minutes after its session went idle. Each new alert gets a read-only review from a Claude Opus 5.5 'Studio Watcher' chat (at most one every 15 minutes, 30 a day), saved to a file.

How you use it

Nothing: a one-line status appears in Anthony's Claude sessions on the Studio. Ask Anthony.

Who has access

Anthony

Watch out: Healthy at 5:44 PM Oct 16: 0 alerts, load 1.2 on 16 cores, disk 59% free. It does not send texts or emails (by Anthony's 'no texts' rule), so alerts are seen only when someone opens Claude on the Studio. It does not yet watch the older crontab jobs (Drive photo intake, lot research, update sender, photo worker) or the stopped invoice sender, which is why the 4-week Drive intake failure is not flagged. Its nightly-pull rule starts Oct 17, and its Monday permit-check and email-routine rules start Oct 19.

What it is

A small command that runs any background helper with a hard limit on processor time and total run time, plus a standing instruction to every Claude session on the Studio to use it.

Why we built it

Prevents another runaway like the Oct 13-16 incident, where leftover test helpers spun for 3 days.

How it works

By default a helper started through it is stopped after 10 minutes of processor time or 1 hour on the clock. The global instructions also tell every Claude session to make helpers quit when their input ends, to clean up processes when a test finishes, and to check for leftovers before finishing.

How you use it

Nothing for the team; Claude sessions on the Studio use it automatically.

Who has access

Anthony and every Claude session on the Studio

Watch out: It only protects helpers started through it. Long-running services are managed separately.

What it is

Each time Anthony types in Claude on the Studio, a hook checks (read-only) whether anyone has replied to the email threads he is waiting on, and Claude opens its answer with one line about replies.

Why we built it

Anthony, Oct 4: 'every time I type in chat make sure to read my email to check if we have responses.' He was waiting on go-live answers from Rich, Daniel, Jim, Natalie, Johanna, Kelsey and Tim.

How it works

At most every 90 seconds it reads the ops server's reply-watch list (threads marked 'waiting') and anthony@'s mail from the team in the last 24 hours. The in-thread reminder emails come from the ops server's reply watch, not from the Studio.

How you use it

Nothing; it shows up in Anthony's Claude replies.

Who has access

Anthony

Watch out: Snoozed at 5:45 PM on Oct 16 until Mon Oct 19 at 7:00 AM ET, after Anthony said everyone is out until Monday; it switches back on by itself. 4 threads are still waiting, all already at the 3-reminder maximum, so no more reminders will go out: Rich x2 (which of the 6 liquidation automation switches to turn on, Oct 1; install Discord, Oct 5), Jim (install Discord, Oct 5) and Johanna (install Discord, Oct 5). At 5:51 PM on Oct 16 Anthony closed four stale threads (the Madden stage-10 date, which has passed; the BuilderTrend stage-photo how-to to Tim; the Sep 25 liquidation overview to Rich; and the ring-list confirmation to Daniel, which Daniel answered in more depth outside the thread), and at 6:07 PM Jacob's Discord thread (he joined Oct 7). On Mon Oct 19 at 7 AM a Studio job emails Rich, Jim, Johanna, Natalie, Kelsey and Alessandra a Discord invite and closes the three old Discord threads, which leaves only Rich's automation-switch question.

What it is

A read-only company-knowledge server on the Studio that Claude apps plug into. It offers knowledge-base search, Drive folder browsing, full documents, live home status from the GB Inventory Sheet, the Systems Guide, and the Closings Log.

Why we built it

One shared toolbox, so every Claude (Anthony's, Tim's, Daniel's) and the ops Assistant see the same company facts.

How it works

It runs around the clock and is reachable only over the company's private network. Each caller has its own key and role (staff, leadership, or owner), and the role decides what it can see. Company login sheets are always withheld.

How you use it

Ask your own Claude. Daniel's Claude is connected with owner-level access, for example 'list this year's home sales from the Closings Log'.

Who has access

Anthony's Studio and Mini Claude, Tim's Claude, Daniel's Claude (owner role), and the ops Assistant

Watch out: Running and used on Oct 16 by Anthony's sessions, Tim's Claude (3:18 PM) and Daniel's Claude (connected at 7:05 AM but made no tool calls; its last calls were Oct 13). Searches can take 15-45 seconds when the Studio is busy. A key for the Discord bot exists but is not connected yet.

What it is

The part of the ops Assistant tab (chat and voice over live company data) that does the thinking. The page lives on ops.gibraltarbuilders.info; the back end runs on the Studio and is reached through a private Cloudflare tunnel.

Why we built it

To let staff ask questions about homes, FD Funding and sync status by chat or voice inside ops.

How it works

Messages from the Assistant tab travel through the tunnel to a small service on the Studio. It re-checks the person's ops login, answers with a cloud model hosted by Groq, and pulls facts from the Studio connector and from ops using that person's own access, so nobody sees more than their login allows. Speech is transcribed by Groq and replies are spoken with an ElevenLabs voice.

How you use it

ops.gibraltarbuilders.info, Assistant tab

Who has access

ops users with the Assistant tab

Watch out: The service and tunnel have run 13 days without a restart; the tunnel reconnected itself briefly at 3:34 AM Oct 16. The last recorded use of company data through it was Oct 5, so it is lightly used. Questions and answers go to Groq and ElevenLabs (cloud), not to the Studio's local models.

What it is

The program that keeps Gibraltar's Discord bot, Hermes, connected and answering. It runs on the Studio around the clock. What the bot can do is covered in the Discord bot section.

Why we built it

Discord is the company's chat, and Hermes is the bot that answers there from company knowledge.

How it works

The system starts it at login and restarts it if it crashes. It holds the Discord connection, searches the knowledge-base snapshot, and (owner-only) can operate the Studio's screen; background clicking, typing, scrolling and app switching were set to 'always allow' on Oct 1, so those steps no longer wait for his Approve button.

How you use it

Discord: mention the bot. It answers anyone in the server (today Daniel, Anthony, Tim and Jacob).

Who has access

Anyone in the Discord server can talk to it; only Anthony can use its owner tools (desktop control, portal status).

Watch out: Up for 4 days without a restart. Its Discord connection dropped and reconnected itself 74 times between Oct 13 and Oct 16 (17 on Oct 16, last at 2:39 PM) while the Studio was overloaded, and it has been quiet since. It searches the Oct 13 knowledge snapshot plus live changes (see the knowledge base entry). It has no scheduled jobs of its own.

What it is

Two small background programs: one stops the Studio from going to sleep, and a menu-bar shield records every remote login (SSH) and screen-sharing session into the machine.

Why we built it

Every job above depends on the Studio being awake. The login monitor was built after a remote login into the Mini on Sep 22 left no trace in any macOS log.

How it works

The keep-awake program blocks idle and system sleep. The shield checks open connections every 2 seconds and logs each session's start, end and length, which shows connections like Tim's bridge and the Mini's checks.

How you use it

Nothing. The shield icon at the top right of the Studio's screen shows live sessions.

Who has access

Anthony

Watch out: Both are running (keep-awake for 76 days). They come from Anthony's Hively tooling but protect the machine that runs Gibraltar's systems. The Studio must still be powered on and logged in as 'rich'; a power cut or restart stops the screen-dependent jobs until someone logs in.

What it is

A free local database on the Studio built from Florida public records: the state tax roll (owners and sales), Brevard property-appraiser data, and Sunbiz company officers. It answers 'who owns this parcel, what else do they own, who is behind this LLC, is the owner absentee'.

Why we built it

To replace owner, parcel and LLC lookups bought from Datafiniti, REAPI and Serper.

How it works

Scripts download the state's public bulk files (the tax roll is published three times a year), load them into a local database, and answer lookups from disk at no cost. It has no phone numbers: a paid skip-trace is still needed, but only for owners already identified.

How you use it

Ask Anthony, or Claude on the Studio, to run a lookup. It is not connected to ops.

Who has access

Claude sessions on the Studio

Watch out: Its own notes say it is not wired into any job, schedule or outreach path. The data was loaded Jul 31 to Aug 1 and has not been refreshed since. The daily lot research reads Brevard's live parcel map directly instead. Unconfirmed whether today's 'accountability sweep on free county data' (being built from Daniel's MacBook Pro, dry run only) uses this database.

What it is

Rich's Mac Studio in the Gibraltar office (64 GB of memory). Anthony does all Gibraltar development and AI work here, signed in to the 'rich' Mac login. It runs about 18 scheduled Gibraltar jobs and 6 timed (cron) jobs.

Why we built it

One always-on machine that can run browser robots, local AI models and nightly jobs, without paying for cloud servers or AI calls.

How it works

Its scheduled jobs include: the photo worker (processes uploaded photos, every 30 s), the stage-update sender, the nightly backup, the daily lot research, the Drive mirror and its every-minute refresh, the knowledge connector, the ops Assistant and its secure tunnel, the Discord bot, Tim's bridge (questions, workers, nightly pull), the closings sync, the BuilderTrend robot, the delay index, Tim's syncs and routines, and a watchdog every 10 minutes. Local AI models (Ollama) run here at no per-use cost. FileVault disk encryption is on with no automatic login, so after a reboot or power cut nothing runs until someone types the Mac's password at the machine. The Mac Mini checks the Studio every 15 minutes to catch exactly that.

How you use it

Nothing: it runs by itself. Daniel reaches its knowledge through his own Claude, and sometimes drives Anthony's Studio Claude session through Remote Control. Anything physical (a restart, typing the password) means asking Anthony or someone in the office.

Who has access

Anthony (day to day, as the 'rich' Mac login). There is a second Mac login named 'daniel' with administrator rights; the notes don't say whose it is.

Watch out: The office router's 'Security Shield' blocks some websites from this machine, including the Datafiniti lot-data service (still blocked today). There is no Time Machine and no external backup disk. Backups go to this Mac and a second copy to the Mini only. The Messages app is signed in to a non-Gibraltar Apple ID. Texting Anthony from it is switched off by his rule.

What it is

Anthony's own Mac Mini (Hively's machine, on his private network). For Gibraltar it is the go-between to the cloud server, the GitHub seat and the second backup copy.

Why we built it

It holds the server administrator key, so the Studio does not have to. It watches the Studio from outside, and Hively's knowledge service lives here.

How it works

1) It is the only machine with full (root) access to the ops server. Studio jobs that send email or edit Google files hop through it. 2) GitHub account work (issues, Actions, billing) is supposed to run here as AnthonyRijos. 3) Hively's knowledge service on it pulls copies of Tim's Claude notes from his MacBook every 15 minutes. 4) It receives a second copy of the nightly backup. 5) Every 15 minutes it checks that the Studio is up and its jobs ran. 6) Anthony's own Claude and a second AI reviewer ('Astra') run here.

How you use it

Nothing for Daniel. Ask Anthony.

Who has access

Anthony. The Studio's Claude has an SSH key into it.

Watch out: Its GitHub command-line login (both accounts) has been invalid since Sep 28 and was still invalid today. Plain git pushes from the Studio still work. Its own copy of the Gibraltar code is stale (last updated Aug 26).

What it is

Anthony's portable dev machines, a MacBook Pro and a MacBook Air, both on his own private network next to the Mini. Daniel also has a MacBook Pro of his own. That is a separate machine and is covered in the 'Daniel's Mac Studio and MacBook Pro' entry.

Why we built it

So Anthony can work away from the office.

How it works

Anthony's MacBook Pro can reach the Studio over the private network. The first 3D walkthrough test was built on the Air on Sep 29. No Gibraltar scheduled jobs run on either. No Mac in the estate has Xcode, which is why the iPhone app is built in the cloud.

How you use it

Nothing for Daniel.

Who has access

Anthony

Watch out: Both were offline when checked today. Do not confuse Anthony's MacBook Pro with Daniel's: the 'macbook-pro' seat in the shared code activity log is Daniel's MacBook Pro. Its commits are signed by Daniel, and it has about 215 entries since Aug 13 (CRM, sales-line phone routing, website), including work at 5:50 PM today.

What it is

A DigitalOcean cloud server named for Hively. It runs Gibraltar's ops dashboard back end (the 'gibralter-api' program) alongside Hively's own services.

Why we built it

It was the server Anthony already had. Moving Gibraltar to its own server is an open census item.

How it works

It holds the ops database (FD Funding ledger, homes and stages, users, reply tracking and more), the Google mail service-account key, the Drive relay and Tim's mail tool, and the Hively close-out census file the Progress tab reads. Besides Hively's back end, the same server also runs another client's services that have nothing to do with Gibraltar. The public web address that reaches the API passes through Hively's Cloudflare tunnel. Full admin (root) sign-in works only from the Mini. The Studio cannot sign in to it directly (its admin key was removed Aug 13); it holds only narrow keys used to fetch the nightly backup.

How you use it

Through ops.gibraltarbuilders.info (sign in); never directly. Changes: ask Anthony.

Who has access

Root: Anthony, via the Mini. Backup-only keys: the Studio's nightly backup.

Watch out: It shares a machine, and a public tunnel, with Hively and with another client's unrelated services. An outage or change on any of them can take Gibraltar's dashboard down. Whether DigitalOcean's own server backups are switched on was not checked.

What it is

Tim Mayes' personal laptop, where his own Claude runs his delay-tracking system.

Why we built it

Tim directs his production work from his own Claude. Sending big jobs to the Studio saves his own Claude usage.

How it works

It is joined to the Studio's private network by a two-way share (Oct 9). His Claude uses three things. The Studio bridge ('gb-studio') asks questions and files build requests. The knowledge connector runs with an owner-level key. And the Mini pulls read-only copies of his Claude notes and transcripts (about 1.1 GB) through a locked-down key that only the Mini can use.

How you use it

Tim uses it himself. Nothing for Daniel.

Who has access

Tim. The Mini has read-only access to his Claude notes folders.

Watch out: The bridge key is tied to the laptop's private-network address. If the laptop is ever re-added to Tailscale, the address changes and the bridge stops until it is updated.

What it is

Daniel's own two computers. His Mac Studio is connected to the office Studio only so his Claude can use the Gibraltar knowledge connector. His MacBook Pro is where his own Claude does build work: it commits Sales CRM, sales-line phone routing and website changes to the same company code repository Anthony uses.

Why we built it

Anthony, Oct 13: 'connect Daniel Casas' computer to the Mac Studio MCP so he can see the closings.' The MacBook Pro is Daniel's own working seat for the CRM and the website.

How it works

The Mac Studio is shared with the office Studio's private network (Oct 13). Daniel's Claude calls the Studio's knowledge connector with his own key ('daniel-claude', owner level). Nothing on the office Studio can sign in to Daniel's Mac. Anthony chose connector-only, not two-way access. The MacBook Pro is not connected to the office Studio. It shows in the shared code activity log as the 'macbook-pro' seat: about 215 entries since Aug 13, roughly 134 of them CRM work and about 26 website work, plus the sales-line ring order. Today it shipped CRM closing-rate tiles, a contact-metrics fix, an 'owner first' ring change on the sales line, and a free-data version of the realtor accountability check (dry run only).

How you use it

Ask your Claude Gibraltar questions (closings, homes, documents). It uses the 'gibraltar' connector automatically. It last connected today at 7:05 AM; its last tool calls were on Oct 13.

Who has access

Daniel

Watch out: If Claude says it can't reach the connector, the usual cause is the office Studio being down or asleep, or the share being removed. Only the Mac Studio has the connector; the MacBook Pro does not. Daniel's seat deploys its own CRM and phone-routing changes (for example today's ring-order change).

What it is

Tailscale is a private, encrypted network between machines. There is no single company network. There are separate Tailscale accounts (the Studio's, Anthony's, Tim's and Daniel's), joined machine by machine with 'shares'.

Why we built it

So Tim's and Daniel's Macs, the Mini and the Studio can talk securely without opening anything to the internet.

How it works

The Studio's account holds the Studio. Anthony's holds the Mini, his laptops and his phone. Tim's holds his MacBook, and Daniel's his Mac Studio. Sharing a machine lets the other account reach it.

How you use it

Nothing day to day. To cut someone off: remove the share in that Tailscale admin console.

Who has access

Studio's account: Anthony (a non-Gibraltar login, not a Gibraltar Google account). Each other account: its owner (Anthony, Tim, Daniel).

Watch out: The access rules between shared machines are among the private-network hardening items Anthony tracks (ask him).

What it is

The gibraltarbuilders.com Google Workspace (Business Starter plan): every company mailbox, Drive and the admin console.

Why we built it

Most Gibraltar data lives in Gmail and the shared Drive, so whoever holds the admin role controls the keys to it.

How it works

There are two super admins: Kelsey (confirmed Sep 17) and rich@ (promoted Sep 17 so Anthony could set up the mail service account). admin@gibraltarbuilders.com is NOT an admin, despite its name. It is an ordinary mailbox, though the GB Inventory Sheet lives in its Drive. Anthony works in the admin console signed in as rich@.

How you use it

Ask Anthony (working as rich@) or Kelsey for anything in the admin console: new mailboxes, password resets, 2-step rules.

Who has access

Super admins: Kelsey, rich@ (used by Anthony).

Watch out: Older notes and one census item still say 'super admin = admin@'. That is wrong. As of Sep 14, 2-step verification was not required company-wide; only four mailboxes had turned it on. Every user sits in one group (organizational unit), so a policy change hits everyone.

What it is

A Google 'service account' (a robot identity), authorised once by a super admin. It can read, send and label mail in any company mailbox, and since Oct 5 can also use Drive. It cannot delete mail.

Why we built it

So the systems can send as the company and read the inboxes they work from, without anyone's password and without breaking when someone changes theirs.

How it works

Google allows it company-wide, but the code holds it to a short list. It may send and label as billing@, scheduling@, kelsey@, anthony@ (the default sender for anything the systems email), bids@ and, since Oct 16, tim@. Since Oct 16 it may also READ any other company inbox, for Tim's requests. It may never send as anyone off the list. The key is kept only on the ops server and is left out of backups.

How you use it

Nothing: systems use it. Emails 'from Anthony' sent by the systems go through it.

Who has access

The ops server. Anthony, who changes the allowed list in code. Tim's bridge requests, through the tim@ mail tool (read any inbox, send only as tim@).

Watch out: Because it is company-wide, the code's allowed list is the real limit. Tim's requests can read every company inbox, including Daniel's. The outside email service (Resend) is unusable: none of its 3 keys has a verified domain.

What it is

A small program on the ops server that edits Google Sheets, Docs and Drive for the Studio, acting as rich@ through the mail robot.

Why we built it

Built Oct 9 for Tim's weekend delay-tracking run. Since Oct 16 it is how Tim's requests edit anything in the shared Drive.

How it works

Callers send it a request and get the answer back; the key never leaves the server. Reads: anything rich@ can see. Writes (since Oct 16): any file or folder in the Gibraltar Builders Shared Drive, plus the GB Inventory Sheet. Always refused: deleting, trashing, changing who a file is shared with, deleting a tab it did not create, writing outside the shared Drive, and writing to Tim's Delay Log 11:40 PM to 12:40 AM. Google's version history keeps every edit undoable. A separate rule files BuilderTrend photos into Drive as scheduling@.

How you use it

Nothing directly. It is used by Tim's requests, the closings sync and Studio Claude sessions.

Who has access

Studio Claude sessions and Tim's bridge workers

Watch out: Edits made this way show in Google as made by rich@, not by the person who asked. The Closings Log sync may only FILL empty cells in the Inventory Sheet, never overwrite them (Anthony's rule). A separate Google Drive connector is also attached to the Studio's Claude. Its recent edits match the relay's, so it appears to be signed in as rich@, but that is not confirmed.

What it is

A read-only Google login (the rclone tool) that copies the whole Gibraltar Shared Drive to the Studio every night. All the AI search tools read that copy.

Why we built it

So every assistant can search company documents locally and for free, without ever being able to change the Drive.

How it works

The nightly copy at about 2:30 AM (10,594 files, 14 GB on Oct 16) is then turned into text and indexed on the Studio. A refresh every minute picks up new files. Its Google permission is 'read only', so it physically cannot write back.

How you use it

Nothing: it feeds the knowledge connector, the Discord bot and the ops Assistant.

Who has access

The Studio

Watch out: It still uses rclone's shared Google app identity, which Google is retiring during 2026. Switching to Gibraltar's own is a census item. No copy was made on Oct 15. The two company password documents in the Drive are copied too, but every search tool excludes them.

What it is

The login that lets the Studio pick up stage photos dropped into Drive and push them to the ops homes/buyer portal. It runs as rich@ through Google's command-line tool (gcloud).

Why we built it

Photos put in a home's 'Stage N' Drive folder should reach the buyer portal automatically.

How it works

Every 15 minutes a Studio job asks gcloud for a Drive token and checks the stage folders.

How you use it

Nothing. Ask Anthony.

Who has access

rich@'s gcloud login on the Studio

Watch out: Every run since Sep 18 has failed with 'could not mint a Drive token', and it was still failing at 5:45 PM today. The quick fix is one sign-in typed at the Studio. The lasting fix is a read-only Drive credential that does not expire. BuilderTrend photos go a different way and are unaffected.

What it is

All Gibraltar code (ops dashboard, CRM, website, voice routing, iPhone app) lives in one private repository in the GitHub organisation 'Gibraltar-Builders'.

Why we built it

One shared, versioned home for the code. Pushing to it automatically redeploys the ops dashboard.

How it works

Members of the organisation: AnthonyRijos and thoughtsinform (Daniel's account), both administrators. The Studio pushes with a repository-only 'deploy key', which can reach this one repository and nothing else (checked today). A push to the main branch redeploys ops through GitHub Actions. The Codemagic app (which builds the iPhone app) is installed on the organisation. Account-level work (issues, Actions logs, billing) is meant to run on the Mini as AnthonyRijos.

How you use it

Ask Anthony.

Who has access

AnthonyRijos and Daniel (thoughtsinform), administrators. The Studio's deploy key (push/pull, one repository). The Codemagic app.

Watch out: The Mini's GitHub logins have been invalid since Sep 28, so no machine can currently use GitHub as a person. Deploys still happen and are checked by looking at the live site. An unpaid bill stopped every Actions run Sep 20-24, until it was paid.

What it is

Two Cloudflare accounts are involved. The 'thoughtsinform' account (Daniel's) holds everything live for Gibraltar. A Hively account holds old leftovers.

Why we built it

Cloudflare serves ops.gibraltarbuilders.info, crm.gibraltarbuilders.info, sales.gibraltarbuilders.info and gibraltarbuilders.com.

How it works

The thoughtsinform account holds: the ops dashboard front end, the CRM and its database, the www-to-main-site redirect, the website (live on Cloudflare Pages since Sep 24), and the secure tunnel through which ops' Assistant tab reaches the Studio. The Hively account holds a dead duplicate of the gibraltarbuilders.info domain and two old site projects. The Studio's Cloudflare logins expire hourly and cannot create API keys. A separate Cloudflare key on the Studio exists only for the nightly CRM database backup.

How you use it

Nothing day to day. Changes: Anthony deploys into Daniel's account; Daniel is the account owner, and his own Claude seat also deploys CRM changes.

Who has access

Daniel (thoughtsinform, owner). Anthony deploys into it, signing in through the Hively login, which can see both accounts.

Watch out: Two domains share the name gibraltarbuilders.info and the dashboard shows them identically; only the thoughtsinform one is live. The API web address still runs through Hively's tunnel. The old WordPress host is kept as a rollback for the website. Who holds the domain-registrar login is not recorded.

What it is

The 'Gibraltar Ops' iPhone app is a wrapper around the ops dashboard. It is built and signed in the cloud by Codemagic, under Anthony's own Apple developer team (the one Hively uses), and handed out through TestFlight.

Why we built it

No Mac here has Xcode, and Anthony's Apple team already had signing set up from his other apps.

How it works

Codemagic pulls the code from GitHub, signs it with a certificate the Apple team shares with Hively, and uploads it to App Store Connect. The first build reached TestFlight on Sep 29. Anthony marked the app complete on Oct 16.

How you use it

Ask Anthony for a TestFlight invite. Today the only tester is Anthony. Tim, Kody and Jacob would first need adding as users on the Apple team.

Who has access

Anthony (Apple team owner, individual membership; his own Codemagic account)

Watch out: TestFlight builds expire after 90 days. This one expires about Dec 28, so rebuild by about Nov 28. The Apple membership renews Mar 14 2027 with no auto-renew card on file. The app lives under Hively's Apple team, not a Gibraltar one.

What it is

Gibraltar's Twilio account ('Thoughts In Form'), with the company toll-free line and the hidden number behind the Liquidation line. The account is Daniel's; Anthony works in it.

Why we built it

Programmable phone numbers for the AI intake line, call routing and texting.

How it works

Liquidation calls forward from RingCentral to a hidden Twilio number (bought Sep 30). It rings the team at once, then hands to the AI agent. Account keys are held on the Studio and the ops server. A separate number borrowed from Hively's Twilio was used before Gibraltar's own.

How you use it

Ask Anthony. Billing and top-ups: Daniel.

Who has access

Daniel (named owner, via his thoughtsinform login). Anthony (console sign-in and API keys on the Studio). The ops server (intake keys).

Watch out: (1) Texts from the toll-free line go out under another business's approved registration, not Gibraltar's. Replies to those texts probably reach nobody. (2) Gibraltar's own texting registration began Oct 5. The brand was pending then, and the campaign step needs a balance top-up and Anthony's OK. (3) The balance was about $15 on Oct 5. At $0 the main line's calls and texts stop. Points 1 to 3 are Oct 5 facts, not rechecked today. (4) The CRM does text through this account: 3 CRM texts were delivered through Twilio between Oct 6 and Oct 9 (checked in last night's CRM backup).

What it is

The company phone system: the main number, the Sales Team queue, the Liquidation Line and staff extensions.

Why we built it

It is the phone system the team already uses. The Liquidation Line forwards into the AI intake.

How it works

The Liquidation Line is its own RingCentral user (extension 113, created by Anthony). It forwards every call to the hidden Twilio number. The main number cannot text. The Sales Team queue line and the Liquidation Line can. In 180 days the account sent zero outbound texts.

How you use it

Normal phone use. Admin changes: ask Anthony or Tim.

Who has access

Tim (Super Admin). Anthony (his own user, which reaches the Admin Portal). Turning on texting for the main number needs Anthony's RingCentral password.

Watch out: Earlier notes list 'RingCentral admin (Jim)' as a login still needed. Who owns the account (billing) is not recorded.

What it is

The AI voice service. Gibraltar uses Daniel's ElevenLabs workspace (Creator plan).

Why we built it

It powers the Liquidation intake phone agent and the spoken replies in the ops Assistant and the Discord bot.

How it works

Gibraltar's systems use a restricted key from Daniel's workspace (it can manage agents, read voices and speak). The intake agent 'Gibraltar Liquidation Intake' lives in Daniel's workspace. The CRM's 'Sales Call Receptionist' agent is also in Daniel's account.

How you use it

Daniel manages the workspace and plan. Changes to agents: ask Anthony.

Who has access

Daniel (workspace owner). Anthony (via the key on the Studio).

Watch out: The key cannot see the credit balance, and one pool of credits pays for the phone agent, the Assistant and the bot. Caps limit the bot's voice use. An older copy of the intake agent, and a borrowed number, still sit in Hively's ElevenLabs/Twilio waiting to be deleted.

What it is

Outside AI services other than Claude and ElevenLabs, plus the free local models on the Studio.

Why we built it

Fast, cheap AI for sorting, transcription and chat. Local models cost nothing per use.

How it works

Groq (an outside AI service) runs the ops Assistant's chat, the bot's voice transcription, and the FD email sorter, which sends parts of closing documents and bank emails to Groq to classify them. Its key is on the ops server, with a copy on the Studio. Local models on the Studio run the Discord bot's brain, the photo-stage guesser and the search index. Rich's personal Grok (xAI) login drives an older bot profile that Gibraltar's bot does not use. It was blocked by a spending limit on Sep 29 but works again (successful calls through Oct 16, mostly non-Gibraltar work); it was used a few times in Aug and Sep to review Gibraltar code.

How you use it

Nothing: used by the systems.

Who has access

Groq key: the ops server and the Studio. Grok: Rich's personal subscription.

Watch out: Whose Groq account it is, and whether its 'don't keep data' setting is on, is not recorded. Grok is deliberately kept out of anything that talks to customers.

What it is

The Claude AI seats that do the work. The Studio's Claude runs under Rich's account; Tim and Daniel use their own Claude accounts.

Why we built it

Claude does the building, the research and Tim's requests.

How it works

The Studio's Claude login runs Anthony's sessions, Tim's worker sessions and the 'Studio Watcher' reviewer. Two connectors are attached to it through claude.ai: JobTread (see JobTread) and a Google Drive connector, which appears to be signed in as rich@. Tim's and Daniel's own Claudes reach company knowledge through the Studio's knowledge connector with their own keys.

How you use it

Daniel: use your own Claude. To act on the Studio, Anthony's session can be driven through Remote Control.

Who has access

Studio seat: Anthony (and Daniel, via Remote Control of Anthony's session). Tim and Daniel: their own seats.

Watch out: Tim's workers stop working if the Studio's Claude login lapses. 'Not logged in' was seen a few times on Oct 14, and workers restart automatically up to 3 times.

What it is

JobTread is where the project managers have kept schedules and tasks since Oct 12, replacing BuilderTrend. The Studio reaches it through a claude.ai connector signed in as Rich Pierce, with read AND write access.

Why we built it

Tim's delay tracking and job-plan work needs to read and update JobTread tasks.

How it works

On Oct 9 (10:23 PM ET) Anthony re-approved Rich's JobTread connector with full access, replacing an earlier read-only grant. Checked today, the grant belongs to Rich Pierce at Gibraltar Builders LLC. Tim also has his own JobTread grant in his own Claude. JobTread can only be reached through the Claude connector, so plain scheduled jobs cannot use it.

How you use it

Tim asks through his bridge. Others ask Anthony.

Who has access

JobTread admins (as of Oct 9): Rich, Tim, Kelsey. The Studio's Claude (as Rich, read and write). Tim's own Claude.

Watch out: Changes made from the Studio show in JobTread as made by Rich Pierce. The brief for this report called it read-only; it is not.

What it is

A robot browser on the Studio, signed in with rich@'s BuilderTrend login. It reads the supers' Daily Log photos and files them into ops (Homes & Stages) and into each home's Drive 'Stage N' folder.

Why we built it

Asked Oct 5: get every home's stage photos out of BuilderTrend into ops and Drive at the same time.

How it works

It uses the web page's own data feed in its own Chrome profile (Anthony signed it in once). A local AI model suggests a stage when the date alone can't tell. Doubtful photos wait in ops for a person to approve. It never deletes anything.

How you use it

See the photos on ops.gibraltarbuilders.info > Homes & Stages. If BuilderTrend logs it out, ask Anthony to sign the robot in again.

Who has access

rich@'s BuilderTrend login (signed in by Anthony)

Watch out: The project managers moved to JobTread on Oct 12, so fewer new BuilderTrend photos are expected. Clicking around inside the robot's window can change BuilderTrend's saved filters.

What it is

Two robot Chrome windows on the Studio, one for the UES inspection portal and one for the Palm Bay iMS permit portal, plus a nightly script meant to feed inspection dates into Tim's Delay Log.

Why we built it

Tim's bonus/delay system counts days between inspections, which live only in these portals.

How it works

Each portal has a window a person signs in to by hand. A keep-alive script refreshes each one every 10 minutes and never types or clicks. Today the windows are read only by the Monday permit check (first real run Mon Oct 19) and by jobs run by hand. The nightly script that would compare the portals with the last read and propose Delay Log rows (never 'Confirmed') is built but not installed. The Discord bot can report the windows' status to Anthony.

How you use it

Nothing day to day. Someone at the Studio must sign a window in by hand when it is signed out, and Tim must accept the UES disclaimer himself. Ask Anthony.

Who has access

UES: Tim's login (he must accept its disclaimer himself). iMS: Gibraltar's contractor login, signed in by hand once and signed out since Oct 13.

Watch out: UES: the session is alive, but the disclaimer blocks reading until Tim accepts it. IMS: signed out since Oct 13 (still so this evening). The windows do not come back by themselves after a Studio restart. The automatic date feed is not installed, so inspection dates do not flow into the Delay Log on their own. Tim reads UES from his own Mac today.

What it is

Paid property-data services behind the Friday sold-homes sweep and the lot finder's owner and distress lookups.

Why we built it

To find off-market lots and check sold comparables.

How it works

Their keys are held on the Studio. The daily county lot research (free) still runs at 6:30 AM. Since Oct 9 the Friday paid sweep has been switched off by Anthony until a spending guard (a confirm step before paid pulls) is built.

How you use it

Lots work happens on ops > Off-Market Lots. Paid pulls: ask Anthony.

Who has access

The Studio

Watch out: Datafiniti is blocked by the office router's 'Security Shield' (checked again today), and its plan ran out of credits in July. Both need fixing before it works. REAPI's credits have lapsed (payment-required errors). Tracerfy was fine as of August.

What it is

A plan to read Gibraltar Builders LLC's QuickBooks Online books (cost of goods, overhead, expenses, interest, cost by project and by month) into the dashboard. It is separate from FD Funding's ledger.

Why we built it

Rich, Sep 16: track cost of goods sold, overhead, expenses and interest, project cost by month and cost by project from Gibraltar's QuickBooks Online.

How it works

Phase one would be read-only. Gibraltar's QuickBooks admin authorises the connection once; the systems then copy the books and match QuickBooks projects to home addresses for review. Nothing would be written to QuickBooks. Today the P&L 2026 tab reads Alessandra's spreadsheet instead. The old QuickBooks Desktop link is dead.

How you use it

Nothing yet.

Who has access

Nobody connected yet. Gibraltar's QuickBooks admin login holder is not recorded here.

Watch out: It needs the QuickBooks admin to authorise it, plus Kelsey's answers: which accounts are job cost, whether homes are set up as Projects, and whether receipts are tagged by home. Separately, invoice forwarding to QuickBooks is broken: 126 vendor invoices queued between Aug 26 and Sep 15 were never sent, and billing@ mail since Sep 15 is not being sorted at all (see the Vendor-invoice pipeline entry).

What it is

The 'Gibraltar Builders' Discord server is the company chat. The 'Gibraltar Hermes' bot runs on the Studio and answers from company knowledge.

Why we built it

Anthony, Sep 30: Discord is the company comms platform and Hermes is its bot.

How it works

The server belongs to Daniel's Discord account. Anthony manages it by being signed in to the Studio's Chrome as Daniel (agreed Sep 30). The bot's settings (changed Oct 8) let anyone in the server use it, and that is the live setting today. Desktop control and the portal status tool stay owner-only, for Anthony's turns alone. Drafting texts and emails is open to any member: the person who asked must tap Send. A fresh #general invite link (it never expires) goes to Rich, Jim, Johanna, Natalie, Kelsey and Alessandra by email on Mon Oct 19 at 7 AM.

How you use it

@mention Hermes in the server, or DM it.

Who has access

Server owner: Daniel's account (also signed in on the Studio's Chrome). Bot owner tools: Anthony. Members today (6): Daniel (owner), Anthony (two accounts), Tim, Jacob and the Hermes bot. Not in the server: Rich, Jim, Johanna, Natalie, Kelsey, Alessandra and Kody.

Watch out: Every channel, #admin included, is visible to every member, and Tim and Jacob are now members. The server does not require 2-step login for moderators. The invite link never expires; Anthony decided to raise the server's verification level and tighten sign-in only after everyone has joined. Today's census still says the bot answers Anthony only, but its live setting answers everyone; the census item is out of date. Financial and closing data is kept out of Discord on purpose.

What it is

Personal email-and-password logins to the ops dashboard, with tabs granted per person, some of them view-only.

Why we built it

Each person sees only the tabs they need. Money tabs can be shared read-only.

How it works

Sign-ins last 30 days. Admins see every tab except Progress, which only rich@, jim@ and daniel@ may open. Accounts on the live server today: - shared 'admin': administrator, 17 active sign-ins, last new one Oct 9. - Daniel: administrator, no sign-in in the last month. - Alessandra: FD Funding, P&L 2026 and Settings; active today. - Tim: administrator since Oct 16; last signed in Sep 25. - Kelsey: E-signatures and AP Comms, plus FD Funding and P&L view-only; active today. - Kody and Jacob: Homes & Stages only; never signed in. - Jim: FD Funding and P&L view-only; never signed in. - rich@: administrator, created Oct 12, signed in today. The Assistant tab is open to everyone signed in. Machines use a separate sync key, not a person's login.

How you use it

Sign in at https://ops.gibraltarbuilders.info. New logins and tab changes: Users & Access tab (admins), or ask Anthony.

Who has access

The 9 accounts listed

Watch out: Jim has never signed in. His 7-day temporary password expired today at about 10:11 AM ET, and a new invite link sent today is good until Oct 23; with it he sets his own password. Tim has been an Administrator since Oct 16, alongside Daniel and rich@. Anthony has no named login and works through the one shared 'admin' account (an open census item). Kody's and Jacob's invite links expired Oct 5 unused. The notes don't say who uses the rich@ login.

What it is

Logins to the sales CRM, a separate system with its own users.

Why we built it

Sales agents work their leads there.

How it works

Last night's backup holds 11 CRM users: 4 admins and 7 agents (some may be inactive). Agents sign in by emailed link. Daniel's two addresses are the built-in admins. Creating sign-in links for new seats needs the CRM admin key, which is not on the Studio.

How you use it

Sign in at https://crm.gibraltarbuilders.info. New seats: Daniel.

Who has access

11 CRM users; Daniel is the admin

Watch out: Seats for Johanna and Chris are an open census item (it also names Natalie, who already has a CRM admin account). The CRM sends queued messages every minute; nobody approves each one. Its texts do go out through Twilio (3 delivered Oct 6 to 9), and outside email is limited. Most CRM building is done from Daniel's own Claude seat.

What it is

A read-only service on the Studio that gives any approved Claude the same company tools: document search and folder browsing, read a document, live home status from the Inventory Sheet, the Systems Guide and the closings log. Each caller has its own key and level.

Why we built it

So every assistant (Tim's, Daniel's, Anthony's, the Assistant tab) shares one set of tools, with one place to grant or revoke.

How it works

Three levels: - staff: public and internal documents; prices and dates stripped from home status. - dev: adds restricted documents and leadership home fields. - owner: everything, including escrow, wire, W-9, HR and payroll documents, unredacted. The company password documents are withheld from every level. Keys: - owner: daniel-claude (Daniel), tim-claude (Tim, owner since Oct 16). - dev: studio-claude, mini-claude, ops-assistant-dev (used for admins in the Assistant tab). - staff: ops-assistant-staff, hermes (made but not connected). - testing: test-staff and test-dev. Each call is logged with the caller's name. Any key can be revoked with one command.

How you use it

Daniel: just ask your Claude. Revoking or adding a key: ask Anthony.

Who has access

The key holders above

Watch out: The two testing keys are still valid (last used Oct 13). Searches can take 15-45 s when the Studio is busy.

What it is

Tim's Claude can ask the Studio's Claude read-only questions and file build requests that worker Claude sessions on the Studio carry out.

Why we built it

Anthony, Oct 12: let Tim's Claude hand whole jobs to the Studio's Claude, so the work runs on the Studio's account.

How it works

Tim's Mac holds an SSH key that works only from his laptop's address and can run only one gatekeeper script. Questions go to a locked-down Studio Claude that can only read a few folders and has no secrets. Requests start a worker in its own session, up to 3 at a time and 200 new requests a day. Since Oct 13 Tim's request IS the approval. Workers still refuse, automatically: moving money, handling credentials, permanent deletes, releasing the shared ops app or CRM, and editing the bridge itself. Every call is logged. A 12:45 AM job files Tim's nightly Delay Log pull as a request.

How you use it

Tim uses 'gb-studio' from his Mac. Others: nothing.

Who has access

Tim (his Mac only)

Watch out: It is switched off by deleting one line on the Studio. 17 requests filed so far, the latest today.

What it is

Every system above needs keys (passwords, API keys, phone and email credentials). This is how they are handled, without saying where any of them are kept.

Why we built it

So nobody has to guess, and so a forwarded email or a lost laptop never exposes a key.

How it works

Keys live only on the machines that use them (the Mac Studio, the ops server and the hosting accounts), in owner-only files or the Mac Keychain, never in documents or chat. Each person who connects their own Claude (Daniel, Tim) holds only their own key, and any key can be revoked on its own without touching the others. Nothing in this report is a key, and logins are never sent by email: people get one-time links and choose their own passwords.

How you use it

Ask Anthony to issue, rotate or revoke any key or login.

Who has access

Anthony holds all of them. Tim and Daniel hold only their own connector or bridge keys.

Watch out: Two company password documents still live in the shared Drive (and so in the Studio's read-only copy of it). Moving those into a password manager is on the open-questions list.

What it is

Every night the Studio pulls a verified copy of the ops database (FD Funding ledger and all), the CRM database, both key Google Sheets, the code history and the server settings. It keeps 30 days and pushes a second copy to the Mini.

Why we built it

The FD ledger carries real money. The CRM's own built-in undo is not a backup.

How it works

The backup runs at 2:30 AM. It checks row counts against the live database and emails on any failure. An 8:00 AM watchdog alerts if the backup didn't run at all. Last night: status OK, 0 failures, finished 3:11 AM, second copy confirmed on the Mini.

How you use it

Nothing: it runs by itself. Restores: ask Anthony.

Who has access

Anthony

Watch out: There is no offsite or cloud copy, and no Time Machine on either Mac. In August a code update once silently replaced the working backup script; it is committed now. The Drive itself is not backed up separately; the Studio's mirror is a read-only copy.

What it is

Pieces of Anthony's other company, Hively, that Gibraltar systems currently depend on.

Why we built it

Anthony built on what he already had, to move fast.

How it works

Things Gibraltar uses from Hively: - The cloud server (Hively's), plus the public address and tunnel that reach the ops API. - The Apple developer team that signs the iPhone app. - The Hively Organizer's close-out census, which the ops Progress tab reads. - Hively's knowledge service on the Mini, which pulls Tim's notes. - An older copy of the intake agent and a borrowed phone number in Hively's ElevenLabs and Twilio, waiting to be deleted. Gmail sending already uses Gibraltar's own Google account.

How you use it

Nothing. Ask Anthony.

Who has access

Anthony

Watch out: Moving the API to a Gibraltar-owned server or tunnel is a 'later' census item.

What it is

Outside marketing accounts the sales programs need.

Why we built it

Lead ads, social posting, local search and MLS listings.

How it works

Meta: the CRM receives Facebook lead-ad leads directly (since Sep 24) through the 'Thoughts In Form' business portfolio. It can sync ads but cannot post. Google Business Profile: no connection, and whose Google account owns it is not recorded. LinkedIn: no connection; who administers the company page is not recorded. MLS: no data feed; Natalie still has no MLS access under Jim's broker.

How you use it

Daniel owns these decisions (census 'Access & credentials').

Who has access

Meta business portfolio: Daniel. Others: not recorded.

Watch out: Census items still open: Meta roles for Natalie and Daniel, Business Profile managers, the LinkedIn path and the MLS feed decision.

What it is

What protects Gibraltar's systems today, and the exposures Anthony has knowingly accepted, one line each.

Why we built it

So leadership knows the trade-offs. Anthony judges risk by what an outsider could do today; insider and foothold issues are kept on a backlog (Oct 4 audit: 0 critical, 7 high, 16 medium, 30 low).

How it works

Protections in place: - Personal ops logins with per-tab and view-only rights. - A separate key for machines. - The mail robot cannot delete mail and is held to a short list. - The Drive relay never deletes, trashes or changes sharing. - The Drive mirror is read-only. - Connector keys are per person and revocable; password documents are withheld. - Tim's bridge key works from one laptop and runs one script. - The Studio disk is encrypted. - Backups are verified nightly, and watchdogs check jobs.

How you use it

Ask Anthony for the full technical list; he tracks the hardening items one by one.

Watch out: Known gaps that need a business decision: (1) the GB Inventory Sheet can be opened by anyone who has its link; (2) Anthony works through one shared "admin" ops login rather than a named one; (3) the Discord server is open to anyone with the invite link and does not require 2-step sign-in (Anthony will tighten it once everyone has joined); (4) Google 2-step verification is not required company-wide; (5) closing-document text is sorted by an outside AI service (Groq); (6) backups stay on our own machines, with no offsite copy. Anthony also tracks a handful of technical hardening items on the Studio and the private network; ask him for that list.

What it is

A single running log inside the ops code repository. Every Claude session on every machine records there what it is about to edit, each commit it makes, and when it is done, so two sessions do not overwrite each other's work.

Why we built it

Several machines and Claude sessions change the same code: Anthony's Mac Studio and Mac Mini, and Daniel's MacBook Pro. Built Aug 13, 2026 as 'one general log any chat can read'.

How it works

A session claims a piece of work before editing and releases it afterwards. A commit hook adds one line per commit. Session start and end hooks record when sessions open and close and release claims a session forgot. A claim older than 12 hours shows as abandoned. The log only ever has lines added, and git merges it without conflicts. A new Mac names itself the first time Claude opens the repo, so no setup is needed.

How you use it

Nothing to do. Daniel's own Claude on his MacBook Pro already writes to it: 213 entries, including today. Ask Claude 'who is editing X' to read it.

Who has access

Every machine and person with access to the code repository

Watch out: Entries so far: 380 from the Studio, 213 from Daniel's MacBook Pro, 38 from the Mini. Known hazard: the claim and release commands push to the shared main branch, and in the Studio's shared copy of the code that can publish another session's unfinished work. Anthony checks for this before pushing.