Files
gbrain/docs/tutorials/company-brain.md
Garry TanandClaude Fable 5 f94b9ef316 docs: repo-wide accuracy + MECE reform from the 9-bucket markdown audit
A code-grounded audit of every markdown file (root, architecture, guides,
mcp, tutorials, docs-root, operations/eval/designs, skills, recipes) followed
by a fix wave with per-bucket ownership. Four classes of change:

Accuracy — every documented command/flag verified against src/ before writing:
dead commands replaced with working ones (pages purge-deleted, jobs watch
--follow, gbrain restore, import-based Obsidian flow, space-separated --scopes,
real thin-client recipes, working isolation verification, real supervisor
restart procedure, curl-based ngrok health check, real minion_mode setter);
count drift fixed with rot-proof phrasing (100+ ops, 50+ bundled skills via
skills/manifest.json, 140+ engine methods, KNOBS_HASH_VERSION pointer instead
of hardcoded versions); stale claims corrected (search-mode defaults, RETRIEVAL
pipeline order incl. autocut, sentinel rules, refusal-list mechanism, engine
snapshot, shard cap 2400s + EXIT-HANG classifier in TESTING.md, latest-stable +
publish-template documented in RELEASING.md as release.yml promises).

MECE — one home per concept, pointers elsewhere: test isolation → TESTING.md;
OAuth registration + --bind/--public-url lore → DEPLOY.md; mode bundles →
guides/search-modes.md (the home the CLAUDE.md dispatcher always promised);
merge contract → schema-packs.md; WAL ladder → ENGINES.md; quiet-hours →
quiet-hours.md; capture taxonomy → entity-detection.md; person-page taxonomy →
compiled-truth.md; brain-first protocol → brain-first-lookup.md; refresh
semantics → refresh-algorithm.md; KEY_FILES.md deduplicated (58 extension
entries merged, one entry per file); infra-layer.md rewritten as a pointer page.

Privacy — placeholder sweep across guides, docs, skills, and recipes per the
iron rule; per-release narration stripped from reference docs (current-state
prose only).

Bootstrap coverage — AGENTS.md pointer, RESOLVER routing row, INSTALL.md path,
tutorial cross-links, keyless-mode sections in spend-controls/headless-install.

skills.lock.json regenerated; llms.txt/llms-full.txt rebuilt. Gates: verify
36/36, typecheck clean, doctor 96/96, skills-integrity + resolver + build-llms
+ config-set + migrations all green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-08-09 17:22:42 -07:00

38 KiB
Raw Permalink Blame History

Tutorial: Extend your personal brain into a company brain

This tutorial picks up where the personal brain tutorial leaves off. You already have a working agent (OpenClaw on Render, talking to you on Telegram, with GBrain as memory and Supabase storing embeddings). Now you want your whole team to use it as shared institutional memory, with each person seeing only what they're allowed to see.

Time: about 90 more minutes on top of the personal-brain install. Cost: under $100 a month sustained for a 25-person company.

If you haven't done the personal-brain install yet, start there first. Come back when you've got the agent responding to you on Telegram. This tutorial assumes that's already working.

I'm Garry Tan. I built GBrain to run my own AI agents at Y Combinator. After a couple of months of multi-user features landing (parallel sync across team sources, per-user OAuth scoping, leak-free isolation across every read path), it's finally usable as a company brain too. This is the recipe I'd run if I were standing it up for a 10-50 person company today.


Part 1: The mental model

What changes when you go from personal to company

The personal brain you built is a single-user system: one git repo, one agent, your stuff. The company brain is the same architecture with three additions:

  1. Multiple sources inside the same brain. Your meeting notes are one source. Each teammate's customer notebook is another. The shared company wiki is a third. They live in the same database but stay independent.
  2. Per-user logins with scopes. Each teammate gets their own OAuth credential. The credential decides which sources they can read and write to. Alice writes to her customer source, reads hers plus the shared one. Bob writes to internal-ops, reads his plus the shared one. Neither can see the other's writes.
  3. Per-person folders, crons, and skills. The shared brain has shared structure, but each teammate gets their own subfolder for their own work, their own scheduled tasks (weekly digest, customer follow-ups), and their own scoped skills.

What this is NOT

It is not a different install. The agent runtime, Supabase backend, GBrain CLI, and AlphaClaw harness from the personal brain stay exactly as you set them up. We're adding to that stack, not replacing it.

It is also not a thin-client-everywhere setup. Your personal agent stays as it is (OpenClaw + Telegram). Each teammate adds their own client of choice (Claude Code, Cursor, Claude Desktop, their own OpenClaw, whatever) and points it at the brain.

What you get that one person's brain doesn't

  • Shared memory. The whole team queries the same brain. The contract notes that Alice wrote on Tuesday show up when Bob asks about that customer on Friday, with citations back to Alice's notes.
  • Scoped privacy. Performance reviews don't leak into customer queries. Legal docs don't leak into sales searches. We fuzz-tested this across every read path and got zero leaks.
  • One sync pipeline. Your brain git repo (or several if you want them isolated per team) feeds the brain. Everyone sees the latest.
  • One operating burden. One server to monitor, not one per user.

Part 2: Switch the brain backend to multi-user Postgres

The personal-brain install uses Supabase as the embeddings layer but the GBrain runtime itself might be using PGLite (single-machine) depending on which path you took. For a company brain, you want a real Postgres for the runtime too. If your personal-brain install is already on Postgres or Supabase end-to-end, skip to Part 3.

If you're on PGLite, migrate:

gbrain migrate --to supabase

This copies every page, chunk, embedding, link, and config over to your Supabase project. Run from the agent host machine, same one you set up in the personal-brain tutorial. Takes a few minutes per 10K pages.

Verify:

gbrain doctor
gbrain stats

Page count and chunk count should match what you had on PGLite.


Part 3: Carve up the brain into sources

The personal brain has one source (called default) holding everything. For a company brain we want multiple. The right shape depends on your org. Here's a typical starting point for a 10-50 person company:

# A shared all-hands source for content everyone reads
gbrain sources add shared --path /srv/brain-repos/shared --name "Shared company wiki"

# A scoped source for sales/customer notes
gbrain sources add customers --path /srv/brain-repos/customers --name "Customer notes"

# A scoped source for internal-only docs (legal, HR, performance, board)
gbrain sources add internal --path /srv/brain-repos/internal --name "Internal-only"

Each --path is a directory on disk where you've checked out a git repo. Create them:

sudo mkdir -p /srv/brain-repos
sudo chown $USER /srv/brain-repos
cd /srv/brain-repos
git clone git@github.com:your-org/shared-wiki.git shared
git clone git@github.com:your-org/customers.git customers
git clone git@github.com:your-org/internal-docs.git internal

You can also keep the existing personal-brain repo as one of the sources. Just pick the role it plays (probably shared if it's already org-wide content). When agents on the host write pages into a source, gbrain sources push <id> (run on the host) commits and pushes those changes back to the source's git repo, so the repo stays the durable system of record.

Two scoping models (pick the one that matches your shape)

There are two ways to scope teammates' access. They suit different deployment shapes.

Model A: separate sources with OAuth scoping (recommended for true multi-user with different AI clients). What this tutorial walks you through. Each teammate gets their own OAuth client, which carries --source + --federated-read flags. The brain refuses cross-source reads at the SQL layer; isolation is database-enforced. Each teammate can run their own MCP-aware client (Claude Code, Cursor, their own OpenClaw, etc.) and the scoping holds.

Model B: one source, directory-based per-person scoping (simpler for one-agent-serves-everyone setups). The shape I actually run in production: a single source called default, with a partners/<slug>/ convention inside it (e.g. partners/alice-example/, partners/bob-example/). Each partner gets their own subdirectory holding their personal pages: partners/alice-example/USER.md, partners/alice-example/concepts/, partners/alice-example/sources/, etc. This is the right model when ONE agent (yours) serves everyone over Telegram or a single shared interface. It's simpler ops, no per-user OAuth. Write scoping within the shared source can be server-enforced: register each per-person client with --bound-slug-prefixes partners/alice-example/ and every slug-mutating write outside that prefix is rejected with permission_denied (v0.42.72.0+). Without the binding, the scoping is convention-only (the agent polices itself). Read scoping stays source-granular in both models — within a shared source, everyone entitled to the source can read every folder.

For most company-brain installs (10+ teammates each with their own AI client), Model A is the right starting point. If you're running the fat-agent-serves-everyone pattern from the personal-brain tutorial, Model B is genuinely simpler. You can also mix: separate sources for the obviously-different ones (customer notes vs internal-only) AND a partners/<slug>/ convention inside the shared source for per-person workspace.

Per-person folder structure inside each source

Inside each source, give each teammate their own subfolder. This is the structure I run:

customers/
├── alice-example/                      ← Alice's customer notebook
│   ├── customers/
│   │   ├── acme-co.md
│   │   └── widget-systems.md
│   └── meetings/
│       └── 2026-05-21-acme-renewal.md
├── bob-example/                        ← Bob's customer notebook
│   └── customers/
│       └── orbit-bio.md
└── shared-customers/                   ← things both can see
    └── all-active-deals.md

Two things this structure buys you:

  1. Each teammate's writes go to their own folder even though they're in the same source. No accidental overwrites.
  2. You can later split a person's folder into its own source (if Alice leaves and a new person takes her accounts, you can move alice-example/ to a new source named after the new person and adjust scoping accordingly).

Same shape for internal/: internal/alice-example/ for her HR docs, internal/bob-example/ for his, internal/legal/ for legal docs everyone can read, etc.

Now sync everything:

gbrain sync --all

Each source syncs in parallel under its own lock so they don't step on each other. Output looks like:

[shared]    100/100 pages
[customers] 240/240 pages
[internal]   85/85 pages
✓ all sources synced

Check the dashboard:

gbrain sources status

You should see all three sources with recent sync timestamps and page counts.


Part 4: Expose the brain over HTTP MCP with OAuth

The personal brain talks to you through the AlphaClaw harness over Telegram. For a company brain we need a path that each teammate's AI client can hit independently. The HTTP MCP server is that path.

gbrain serve --http --port 3131 --bind 0.0.0.0

The --bind 0.0.0.0 is important. By default the server binds to localhost only, which is correct for a personal install but blocks remote teammates. Setting 0.0.0.0 accepts connections from any interface. (Full detail on --bind / --public-url, including the ECONNREFUSED failure mode they prevent, lives in DEPLOY.md — Expose the server.)

The server prints an admin bootstrap token to stderr on first start when run in an interactive terminal. Save it. You'll use it once for the admin dashboard. On a non-TTY start (systemd, Docker, piped logs) the token is hidden from logs — set GBRAIN_ADMIN_BOOTSTRAP_TOKEN yourself or pass --print-admin-token on a trusted terminal instead.

For development, tunnel the local server out via ngrok:

ngrok http 3131 --domain your-brain.ngrok.app

For production, put your server behind a real hostname with a real TLS certificate. Let's call your final URL https://brain.acme-co.com for the rest of this tutorial.

Re-run the server with the public URL so the OAuth discovery metadata matches what clients hit:

gbrain serve --http --port 3131 --bind 0.0.0.0 --public-url https://brain.acme-co.com

You should be able to hit https://brain.acme-co.com/health and get {"status":"ok"} back.


Part 5: Register one OAuth client per teammate

Each teammate (or each AI agent for a teammate) gets their own OAuth client. The client controls what they can write and what they can read.

# Alice (sales): writes customers/alice-example, reads customers + shared
gbrain auth register-client alice-example \
  --grant-types client_credentials \
  --scopes "read write" \
  --source customers \
  --federated-read customers,shared

# Bob (ops): writes internal/bob-example, reads internal + shared
gbrain auth register-client bob-example \
  --grant-types client_credentials \
  --scopes "read write" \
  --source internal \
  --federated-read internal,shared

# Carol (legal): writes shared/legal, reads all three
gbrain auth register-client carol-example \
  --grant-types client_credentials \
  --scopes "read write" \
  --source shared \
  --federated-read shared,customers,internal

Each register-client command prints a client_id and a client_secret. Save both for each teammate. They go into the teammate's local agent config. (The full registration reference — grant types, the /admin dashboard flow, DCR — lives in DEPLOY.md — Register OAuth clients. What follows is the multi-user delta.)

A note on the flags:

  • --scopes "read write" (space-separated, quoted — the OAuth wire format; a comma-separated list is rejected at registration) lets the client query the brain and write new pages. You can omit write for read-only clients (executive summaries, dashboards). The admin scope is needed for operational commands like gbrain remote doctor and is usually reserved for your own admin client.
  • --source controls write authority. A client can only write to one source. Within that source, your folder convention from Part 3 keeps each person's writes in their own subfolder — and you can make that server-enforced with --bound-slug-prefixes alice-example/ (v0.42.72.0+): every slug-mutating write op (put_page, delete_page, tags, links, timeline, revert, raw data) outside the bound prefixes is rejected with permission_denied. Update the binding later with gbrain auth rescope-client <id> --bound-slug-prefixes <p1,p2|none>. Adding a binding to an existing client narrows it in ways you should expect: ops that write by something other than a slug (extract_entities, extract_facts, forget_fact, ontology_propose, sources_add/sources_remove) and POST /ingest become unavailable to that client, and put_page's automatic fact extraction is skipped — all because none of them can be confined to a prefix. Reads are unaffected. See the qm-harness guide for the full model.
  • --federated-read controls read scope. A client can read from one or more sources.

Verify the scoping actually scopes

Before you hand the brain to teammates, verify isolation. The clean way is a thin-client install on a second machine (or a scratch shell): gbrain init --mcp-only writes a config that routes every CLI command through your remote server as one specific OAuth client, so a plain gbrain search exercises exactly the path teammates will use.

# As Alice (on a machine that is NOT the brain host)
gbrain init --mcp-only \
  --issuer-url https://brain.acme-co.com \
  --mcp-url https://brain.acme-co.com/mcp \
  --oauth-client-id <Alice's client_id> \
  --oauth-client-secret <Alice's client_secret>

gbrain whoami                          # confirms which client you're acting as
gbrain search "performance review"

Alice should see results only from customers and shared. The performance-review notes live in internal, which she's not scoped to read. She shouldn't see them.

Now re-run the same check as Bob. On the same test machine, swap the credentials with --force (it overwrites the thin-client config):

gbrain init --mcp-only --force \
  --issuer-url https://brain.acme-co.com \
  --mcp-url https://brain.acme-co.com/mcp \
  --oauth-client-id <Bob's client_id> \
  --oauth-client-secret <Bob's client_secret>

gbrain whoami
gbrain search "performance review"

Bob should see the performance-review notes from internal, plus anything related from shared. He shouldn't see anything that lives only in customers.

If both queries return correctly scoped results, isolation is working. (There is no per-query "act as client X" flag — the thin-client config decides which credential the CLI uses; only the client secret can be overridden at call time via GBRAIN_REMOTE_CLIENT_SECRET.)


Part 6: Set up per-person crons

The personal-brain install runs the dream cycle (overnight enrichment) once per night for one user. A company brain needs per-person crons because each teammate has their own context: Alice wants a 7am customer-pipeline digest, Bob wants a 9am ops-status report, Carol wants a contract-compliance check every Monday.

Each cron is just a scheduled gbrain agent run call scoped to the teammate's client credentials. The schedule lives in the workspace repo (the one AlphaClaw deployed in the personal-brain tutorial), in a crons/ directory. A typical layout:

your-org/myagent/
└── crons/
    ├── alice-example/
    │   └── 07am-customer-digest.md
    ├── bob-example/
    │   └── 09am-ops-status.md
    └── carol-example/
        └── monday-contract-compliance.md

Each cron file declares its schedule and the prompt that the agent runs:

---
schedule: "0 7 * * *"
client: alice-example
---

# Customer pipeline digest

Pull every customer page in customers/alice-example/ that had activity in
the last 7 days. For each, summarize what changed and what the next action
is. Output as a markdown digest, post to Slack #alice-customers, save a
copy to customers/alice-example/digests/YYYY-MM-DD-pipeline.md.

The client: field tells the cron runner which OAuth client to use, which enforces the scoping. Alice's cron can only read Alice's sources and write to Alice's folder. It cannot accidentally touch Bob's customer notes.

To install the cron schedule, commit the file to the workspace repo and let AlphaClaw pick it up on next deploy. The cron-scheduler skill (one of the bundled skills GBrain installed) handles the dispatch. (The client: frontmatter field is a workspace/harness convention — your cron runner reads it and picks the matching OAuth credential; GBrain enforces the scoping once the credential is used.)


Part 7: Add per-person skills

The bundled skills GBrain installs are generic. Your team probably wants a few that are specific to them. Examples:

  • onboarding-new-hire. Only Carol (HR) runs this. Walks through generating a welcome packet, scheduling intro meetings, provisioning accounts.
  • customer-success-followup. Only Alice (sales) runs this. Pulls latest customer page, drafts a follow-up email, posts to her review queue.
  • weekly-team-digest. Only you (admin) run this. Aggregates everyone's published pages into one weekly summary.

Skills are just markdown files in the workspace repo's skills/ directory. The shape:

your-org/myagent/
└── skills/
    ├── onboarding-new-hire/
    │   └── SKILL.md
    ├── customer-success-followup/
    │   └── SKILL.md
    └── weekly-team-digest/
        └── SKILL.md

Each SKILL.md declares the trigger (verbs in plain English the agent listens for) and the procedure. Use the gbrain skillify scaffold <name> command to generate the boilerplate:

gbrain skillify scaffold onboarding-new-hire

That creates the directory + SKILL.md + routing entry. Edit the SKILL.md to describe the procedure, commit, deploy. The agent picks up the new skill on next request.

Per-person scoping for skills is a routing-layer convention, enforced by your agent harness, not by GBrain: declare something like allowed_clients: [carol-example] in the skill's frontmatter and instruct your agent (in its routing rules) to refuse the skill for anyone else. The hard guarantee stays at the data layer — even if the agent runs the skill anyway, Alice's OAuth credential still can't read or write outside her scoped sources.

Shared rule files at the skills root

Alongside individual skill directories, drop a few flat _*-rules.md files at the root of skills/. These are conventions that EVERY skill reads. The ones I run in production:

  • _brain-filing-rules.md. the iron-rule decision tree for "where does this new page belong?" Numbered first-match-wins rules (people go in people/, companies in companies/, meetings in meetings/, etc.). Every ingest skill consults this before creating a page.
  • _output-rules.md. output quality standards (deterministic links built from API data not LLM-composed strings, exact-phrasing requirements for citations, no AI-slop vocabulary).
  • _excluded-people.md. a privacy gate. Names that must never be referenced or attributed in the brain even if they appear in source material. Re-attribute or discard. This is the file that prevents your agent from accidentally publishing things about people you've decided aren't fair game.
  • _operating-rules.md. operational conventions (when to write to brain vs scratchpad, when to ask for confirmation, when to fire a notification).
  • _x-ingestion-rules.md, _x-api-rules.md. per-source rules for specific integrations (Twitter, in this case).

These files turn into the de facto company policy for the agent. Edit one, and every skill that reads it picks up the new rule on the next request. Versioned in git, reviewable in PR.


Part 8: Wire Slack carefully

Slack is the integration most teams want first, and it has enough sharp edges to deserve its own callout. The conventions I run:

Two crons, two jobs. One scan cron that runs every 5-15 minutes and surfaces signals (new threads in channels you care about, mentions of your teammates, decisions). One archive cron that runs nightly and stores the full conversation history. Splitting them this way means urgent signals get acted on fast while the slow archive work doesn't crowd the live channel.

Channel-to-task-ID mapping. Don't have your agent reference Slack channels by their actual channel IDs (C03A8...). Build a topic-registry.json (or similar) that maps each channel ID to a friendly task name (acme-co-customer-success, engineering-standup). Crons and skills reference channels by friendly name; the registry translates to IDs at runtime. This is the file you edit when a channel gets renamed or replaced.

Deterministic links only. When your agent writes a brain page that cites a Slack message, the link MUST be built from API data (workspace ID + channel ID + message timestamp), never composed by the LLM. LLMs hallucinate Slack URLs constantly. The convention lives in _output-rules.md; every skill that touches Slack inherits it.

Dismissed-items state. The scan cron remembers what it has already surfaced. If a channel had a thread on Tuesday that turned out to be noise, the dismissed-items file records it so the Wednesday scan doesn't surface it again. Without this, re-scans become a flood of repeat signals.

Per-channel scoping mirrors per-person scoping. Sensitive channels (#executive, #legal, #performance) should be scoped to teammates with the appropriate --federated-read. The brain stores everything, but who can query for it is gated by the same OAuth client model from Part 5.

The actual skills that implement this in production are named slack, slack-scan, slack-archive. Scaffold equivalents in your workspace with gbrain skillify scaffold slack-scan, then edit the generated SKILL.md to declare your channel mapping and triggers.


Part 9: Onboard each teammate yourself (the botmaster pattern)

This is the part that decides whether your company brain actually gets adopted or sits unused.

Do not just hand a new teammate their OAuth credential and tell them to "try it out." They'll send one query, get a result that doesn't feel personal yet (because their slice is empty), conclude it's not useful, and never come back.

What works instead: I personally onboard each new teammate myself. The flow looks like this.

Step 1: Pre-populate their slice

Before they ever log in, I seed their partners/<their-slug>/ directory (or their dedicated source) with the context they need to feel like the brain already knows them:

  • partners/alice-example/USER.md. a one-page profile: role, focus areas, current top 3 priorities, the kind of questions they tend to ask, the kind of writing they prefer (terse vs detailed, casual vs formal).
  • partners/alice-example/concepts/. 5-10 frameworks or recurring themes that are specifically THEIRS. If Alice runs sales, that's "pipeline stage definitions," "ICP criteria," "objection-handling playbooks."
  • partners/alice-example/sources/. links to the documents they care about (their team's shared docs, their inbox conventions, the dashboards they check).
  • 2-3 example brain entries that demonstrate the shape: a customer page they'd recognize, a meeting note from a recent meeting they attended, an idea they've shared with the team.

Takes me maybe 20 minutes per teammate. The payoff: the moment they run their first query, the brain answers with their context, not a generic response. That's the difference between "this is a cool tool" and "this knows me."

Step 2: Walk them through 2-3 wow flows

Before letting them DM the agent freely, I personally walk them through 2-3 specific flows that I know will land:

  1. A query that demonstrates synthesis: "ask the brain about [a customer they know well]. Notice how it pulls together pages from three sources into one answer with citations." This shows the brain layer in action.
  2. A query that demonstrates gap analysis: "ask the brain about [something it doesn't know yet]. Notice how it tells you what's missing instead of making it up." This builds trust.
  3. A write-back flow: "tell the brain about [a meeting they just had]. Notice how it auto-files, links to the other people who were there, and surfaces related history." This shows the agent's value as a capture tool, not just a query tool.

These three flows take maybe 15 minutes total. By the end, the teammate has seen the brain do something they couldn't have done themselves in that time. They feel powerful.

Step 3: Graduate to DM only after the wow moment lands

After the walkthrough, I give them their OAuth credential and the agent's DM (Telegram, Slack DM, whatever your interface is). I explicitly say "now you can ask it anything, write to it anytime, and it'll keep learning from you."

The order matters. If you give them DM access first and expect them to discover the wow moments themselves, most won't. They'll send one generic query, get a generic answer, and bounce. The botmaster pattern (pre-populate → walk through → graduate to DM) flips the conversion rate.

Repeat this flow for every new teammate. About 45 minutes per person, total. Compared to the cost of an unadopted internal tool, it's the best 45 minutes you'll spend.


Part 10: Connect each teammate's AI client

Each teammate runs their AI client (Claude Code, Codex, Claude Desktop, OpenClaw, Hermes, whatever) configured to point at your brain server. Two pieces, both direct-to-server — there is no local relay in between:

1. The GBrain CLI, as a thin client (recommended for everyone). On their machine:

curl -fsSL https://bun.sh/install | bash
bun install -g github:garrytan/gbrain

gbrain init --mcp-only \
  --issuer-url https://brain.acme-co.com \
  --mcp-url https://brain.acme-co.com/mcp \
  --oauth-client-id <their client_id> \
  --oauth-client-secret <their client_secret>

The thin-client install creates a local config that knows how to talk to your brain but never opens its own database. From then on, plain CLI commands (gbrain search, gbrain query, gbrain think, gbrain whoami, ...) route through your remote server transparently, as that teammate's OAuth client. Local-only commands (gbrain sync, gbrain serve, gbrain embed, ...) are refused with a hint — those run on the brain host, not on teammate laptops.

2. Their AI client, connected directly to https://brain.acme-co.com/mcp. Each client has its own connection shape; the per-client pages in docs/mcp/ are the reference:

  • Claude Code / Codex — one command from anywhere gbrain is installed: gbrain connect https://brain.acme-co.com/mcp --token <token> --install (see CLAUDE_CODE.md / CODEX.md). Note the credential type: gbrain connect for these two agents uses bearer tokens (gbrain auth create <name>), which are full-access. That's fine for you as the admin; for source-scoped teammates, the scoped credential is their OAuth client — use it via the thin-client CLI above and the OAuth-capable clients below.
  • Claude Desktop — remote servers are added through the GUI: Settings > Integrations, URL https://brain.acme-co.com/mcp. Do not put a remote server in claude_desktop_config.json; that file only works for local stdio servers and fails silently for remote ones. See CLAUDE_DESKTOP.md.
  • ChatGPT (CHATGPT.md) and Perplexity (PERPLEXITY.md) — both speak OAuth to the server directly, so per-teammate scoping carries into those tools. Perplexity uses the same client_credentials clients you registered in Part 5. ChatGPT needs an authorization_code (PKCE) client — register one per teammate with the same --source / --federated-read flags.
  • OpenClaw / Hermes forks — if the teammate's own agent runs on a machine with a full local gbrain install, it can use local stdio (gbrain serve) against its own brain and reach yours over HTTP MCP like any other remote client.

Part 11: First real query as a teammate

Have Alice run a real query from her machine. The interesting verb is gbrain think, which gives back a synthesized answer instead of raw pages.

gbrain think "What's the latest update from acme-co? When did we last talk to them?"

What Alice gets back, assuming the brain has been syncing for a week and her sources contain a customer page for acme-co and several meeting notes:

## Answer

The most recent customer contact with acme-co was a renewal-discussion
meeting on 2026-05-18, attended by alice-example and acme-co's CTO. Key
points discussed [customers/alice-example/meetings/2026-05-18-acme-renewal]:

- They are upgrading their plan from team to enterprise.
- Annual contract value is moving from $48K to $180K.
- Decision driver: a new compliance requirement they have to meet by Q3.

Prior contact was a quarterly check-in on 2026-04-03 [customers/alice-example/meetings/2026-04-03-acme-q2-checkin].

**Gap noted:** No customer-success notes have been filed since the
2026-05-18 renewal meeting. If a follow-up has happened, it's not in
the brain yet.

Three things to notice:

  1. Sourced. Every claim cites the meeting note it came from.
  2. Synthesized. Alice didn't read three pages and stitch them together. The brain did.
  3. Honest about gaps. The brain knows what it doesn't know and says so, instead of inventing a follow-up that didn't happen.

That last part is the gap analysis. It's the part of the brain layer that nobody else ships.

Bob asking the same question would get nothing about acme-co. He's not scoped to read customers. He'd see his own internal-ops content if he asked something relevant to that. Carol asking would see both, because she's scoped to read all three sources.


Part 12: Operating the company brain

Three commands do most of the operational work.

Background daemon: gbrain autopilot

The personal-brain install already turned this on. For a company brain, the same autopilot covers all your sources because they live in one database. It runs every five minutes; on a healthy brain (health score 95+) it sleeps; on a brain that's drifting it submits targeted maintenance jobs.

Self-healing: gbrain doctor --remediate

gbrain doctor --remediate --yes --target-score 90 --max-usd 5

Computes a dependency-ordered plan of maintenance jobs that would raise the brain's health score to the --target-score, runs the plan, refuses to spend past the --max-usd cap. Safe to cron.

Monitoring: gbrain sources status and the admin dashboard

gbrain sources status

Returns a per-source dashboard: when each source last synced, how many pages, how many embedded, how many unacked sync failures. The at-a-glance health check.

The admin dashboard at https://brain.acme-co.com/admin shows live request volume, registered OAuth clients, recent activity, and brain stats. Use the admin bootstrap token from Part 4 to log in the first time, then register additional admin users from inside the dashboard.

If agents run as containers on the same Docker host

OAuth source scoping only guards the HTTP MCP path. If the brain's Postgres and your teammates' agent runtimes are containers on the same Docker host, make sure the agents can't reach Postgres directly over Docker's default bridge network — a direct DB session skips OAuth entirely. Put Postgres on its own user-defined network, publish it loopback-only if at all, and never hand agent containers a DATABASE_URL. The copy-paste operator checklist lives in docs/mcp/DEPLOY.md — Co-located Docker workloads.


Part 13: Cost and speed expectations

Real numbers from the published benchmark, running the default stack (GBrain with ZeroEntropy for embedding + reranker):

  • Embedding cost: $0.05 per million tokens. For comparison, GBrain configured with OpenAI is $0.13 (2.6× more expensive), Voyage is $0.18 (3.6× more).
  • Ingest speed: about 22 seconds for a small test corpus of 164 pages on the host machine. For a 10K-page corpus, expect about 20 minutes the first time, then most syncs are incremental and finish in seconds.
  • Query latency: about 122 ms median for a gbrain search. For comparison, the same query through GBrain with OpenAI takes about 282 ms.
  • Synthesized-answer latency: a few seconds, dominated by the Anthropic API.
  • Retrieval quality: on the public LongMemEval benchmark, GBrain hits 97.60% recall at the top 5 retrieved sessions, beating the previous published state of the art at 96.6%. On the in-house BrainBench corpus of relational queries, GBrain beats commodity vector retrieval by 38 percentage points, because the graph layer surfaces relationships that vector similarity alone misses.

Full methodology and per-run receipt JSONs live in the gbrain-evals repo.

For a 25-person company at sustained use, expect about $35 a month in embeddings (ZeroEntropy at $0.05/million tokens), $50 a month in Anthropic calls for the synthesized-answer queries, plus your hosting bill. Under $100 a month for the AI side at most companies your size.


Part 14: Common gotchas

"My teammate can't see anything"

Check gbrain auth list on the host and confirm their client has --source set to a source that actually exists. Empty or null --source means the client falls through to the default source, which probably has no content if you set up three named sources.

"Sync is slow and feels stuck"

The first sync embeds every page, which takes time. Check gbrain sources status for the live page count. If it's climbing you're not stuck, you're just embedding. If you've got a 10K-page corpus and ZeroEntropy is being throttled, the per-source parallel sync looks like progress on three sources at once rather than one source moving fast.

"I see a page I shouldn't see"

This shouldn't happen, but if you suspect it, run gbrain search "<query>" --json from a thin-client install configured with the constrained client's credentials (the Part 5 verification setup) and inspect the source_id field on every returned result. Every row should be in the client's --federated-read set. If one isn't, file an issue with the exact slug and source IDs.

"The synthesized answer is wrong"

The brain layer is grounded in the retrieved pages. If the retrieved pages contain bad information, the answer will too. The gap-analysis note often catches this: if the answer says "based on retrieved pages from date X" and date X is six months ago, the brain is telling you the information is stale. Run gbrain sync --all to refresh and try again.

"OAuth /token endpoint returns 401 for my client"

Verify the client secret matches what was printed at register-client time. The server stores only a SHA-256 hash; if you lost the original, you have to revoke the client and re-register. Use gbrain auth revoke-client <client_id> and re-run register-client.

"Postgres connection is exhausting"

Each parallel sync worker opens its own pool. With three sources and the default four workers per source, you can hit your Postgres connection limit if it's set low. Either reduce the worker count with gbrain sync --all --parallel 2 --workers 2, or raise your Postgres max_connections to at least 100. Supabase's free tier defaults to 60, which is tight.

"I want to add a fourth teammate but they need access to all three sources"

gbrain auth register-client diana-example \
  --grant-types client_credentials \
  --scopes "read write" \
  --source shared \
  --federated-read shared,customers,internal

That's it. Add or rotate teammates as the org grows.


What you built

You now have the personal-brain agent from the previous tutorial, plus a multi-user shared layer on top: three federated sources holding shared, customer, and internal-only content; per-person folders inside each source so teammates' writes don't collide; per-person OAuth clients with scoped read and write; per-person crons that run on each teammate's own schedule with their own scoping; per-person skills the agent only runs for the right person. Each teammate queries the brain in plain English through their AI client and gets back synthesized, sourced answers that are correctly scoped.

What to do next:

  • Wire ingestion from external systems (Granola, Linear, Slack) using the ingestion source contract. Most companies want their meetings auto-ingested so the brain stays current without anyone typing notes.
  • Set up team-specific dashboards through the admin UI. Each team lead can have their own view of brain health and activity.
  • Explore the rest of the brain layer. gbrain whoknows (find the expert on a topic), gbrain find-trajectory (how a metric changed over time), gbrain founder scorecard (especially useful for VC and ops teams), the contradiction-detection cycle that surfaces conflicts between different people's notes.

If you're building in this space (which YC has flagged as the company-brain category in its Request for Startups), you might as well build on this. Everything described above is open source, MIT licensed, and what I run in production behind my own AI agents.

Questions, gotchas, or wins worth sharing? Open an issue at github.com/garrytan/gbrain.