ORGAN 02 · COLLECTIVE

Team Brain

One brain every workspace feeds. Knowledge, decisions, and deliverables in a single queryable store that gets sharper with every push.

Setup your team brain
The AIOS Team Brain home: ask the shared brain in plain English, with KPIs, knowledge growth, and a live activity feed
HOW IT WORKS

One Team Brain

Every user and data source feeds a single shared brain: knowledge, memory, reasoning, and policy in one place. A context engine finds the patterns; the brain surfaces the work and acts on it, getting sharper with every use.

Users on Claude Code, Cursor and the AIOS Workstation send daily updates, and data streams from Slack, Notion, Linear, GitHub and Google are ingested, into one Team Brain holding knowledge, memory, reasoning and policy. A context engine beneath it finds patterns and closes the loop. The brain surfaces a kanban board, a KPI dashboard and a query interface, and takes policy-governed actions: spawning agents, executing tasks, and improving the team's shared harness of skills, hooks and guards.

USERS
Alex Claude Code
Sam Conductor
Jordan Cursor
INTEGRATIONS
+ more
Team Brain Knowledge · Memory Reasoning · Policy · Insights
Context Engine manages context · finds patterns shares insights
SURFACES
ACTIONS
Spawn agentsresearch · outreach · drafting
Execute taskspull from kanban autonomously
Improve harnessa goal → shared skills, hooks & guards
The context layer

We make sense of it.

Everything the team produces feeds one place. We turn it into a clear, per-person timeline of who did what and when, then build a context layer on top that understands the narrative arcs of the work and the gaps: what is moving, what has stalled, and where attention has drifted. All of it queryable in plain English.

Everything the team produces, chat, docs, pull requests, issues, and meetings, is ingested into one place. AIOS normalizes it into a clear, per-person timeline of who did what and when. A context layer sits on top of that timeline and understands the narrative arcs of the work and the gaps: what is moving, what has stalled, and where attention has drifted. All of it is queryable in plain English through an LLM.

Attribution

Nothing it writes is unsourced.

Open any sentence the brain wrote and it comes apart into the items it was built from. Open those and you land on the records they came from. The ingestors normalise what your tools emit but keep the identifier of the original rather than discarding it, so the link survives every step between a raw commit and a claim on a dashboard.

A sentence from a narrative arc, opened downward into what it was built from. The arc reads that routing is wired for the three text jobs, the fallback path is tested, and the work left is the cost ceiling. Opening that claim shows the three attributed items behind it: two GitHub commits and a Linear issue. Opening those shows the source of record for each one, the actual commit URLs and the issue it came from. Nothing the brain writes is unsourced, because the identifier of the original is kept at every step rather than discarded.

THE ANSWER

You do not have to guess. It scores every repo you have.

The brain reads each codebase and grades how agentic it actually is: whether agents have instructions to follow, whether there are evals and tests to check their work, how much of the recent history they wrote, and how ready the repo is for them to work in unsupervised. Then it shows you which parts are missing.

acme/acme-platform TypeScript · Python · Shell last 30 days

AGENTIC SCORE

provisional heuristic · a transformation signal, not ground truth

  • Test coverage 62.4 · 25%
  • Agent scaffolding 90 · 25%
  • Skill breadth 80 · 20%
  • AI-assisted commits 58.4 · 15%
  • Commit cadence 100 · 15%
  • CLAUDE.md
  • AGENTS.md
  • 7 skills
  • 2 commands
  • 61.2% lines
  • 66.4% functions
  • 55.1% branches

AGENT READINESS

L4 83.33%

  • Docs 2/3
  • Evals 1/1
  • Dev env 1/1
  • Testing 2/2
  • Security 1/1
  • Code quality 2/3
  • Observability 0/1
  • Build validation 2/2
  • Agent instructions 4/4
Every repository the team connects, scored on the same scale.
Use cases

Put the shared brain to work.

Three everyday ways a team brain takes work off your plate.

Catch up in one question

Ask “what happened while I was out?” and get a synthesized answer across everyone’s work, instead of a morning of catch-up calls.

See the gaps before they cost you

Duplicated effort, a thread gone quiet, attention drifting from what mattered. The brain surfaces it the moment it appears, not at the next retro.

Point your agents at real context

Agents draft, answer, and act against the same ground truth the team sees, and write their results back, so your automation stays grounded.

What it's built on

The same values as the Workspace.

P1

Modular

Each organ system is independently deployable. Adopt what you need: workspace only, brain only, or the full loop.

P2

Open source

MIT-licensed and self-hosted. Clone the workspace per person, run the brain on your Postgres. No cloud lock-in.

P3

Permissioned

Every action is policy-governed. Teams define what runs autonomously and what requires human approval.

P4

Privacy-first architecture

Individual agent data stays personal by default. The collective brain receives only what members choose to push.

Getting started one click

Deploy the whole brain in one click.

The official Railway template creates Team Brain, Postgres, Graphiti and Neo4j together, generates the required secrets, and shows the resources and charges before anything is deployed.

Deploy Team Brain on Railway

Before you deploy

  • Railway Hobby plan. The $5 monthly minimum includes $5 of resource usage; additional usage is billed separately. An expired trial cannot deploy the template. View Railway plans.
  • Team and first-admin details. You enter these during setup. The admin password must be at least 10 characters: the form accepts a shorter one and the deployment then fails while creating the admin.
  • One provider API key for LLM answers. Bring a key from Anthropic, OpenAI, or OpenRouter. The core deployment works without one, but the Team Brain cannot answer questions until a key is added.

What Railway sets up

  • One Team Brain app. A shared instance for your team.
  • One managed Postgres database. Railway wires it to the app.
  • Neo4j and Graphiti, the context engine. Both run on the private network. Graphiti uses the provider and extraction model you configure in Team Brain Admin, so it needs no separate provider key.
  • Generated application secrets. The template creates them and bootstraps your first admin without printing the supplied password.

No GitHub fork or local Railway CLI is required.

What the template does not do

It does not connect a personal workspace or push any content. Railway shows the services and estimated charges before deployment.

Prefer another host or a local install? The full setup guide keeps the portable manual path.

Three things are required and carry the whole product: any host running Node 20 or newer, the Next.js app, and one Postgres 16 database with the pgvector extension for search. Two further services are optional and can be added later or never. A Python sidecar, run on your own infrastructure, pulls from Notion, Google Drive, Confluence, RSS, web pages and local files. Neo4j with Graphiti is the context engine, which powers the narrative arcs on Pulse. Without either optional branch, the rest of the product is unaffected.

Required: it will not run without these

Required 01

A host

Anything that builds a Next.js app, on Node 20+.

Why we suggest Railway

It detects the app automatically and the official template includes Postgres, Graphiti and Neo4j with private-network wiring. Nothing in the code depends on it: Fly, Render or your own box work the same.

Required 02

Postgres 16

The source of truth. One database per team instance.

What lives in it

Every item, task, decision and attribution. The brain is Postgres-native: the timeline, search and admin surfaces all read it directly.

Required 03

Two secrets

AUTH_SECRET signs sessions, SECRETS_KEY encrypts connector tokens.

# AUTH_SECRET, then SECRETS_KEY
openssl rand -base64 32
openssl rand -base64 32

Needed for the brain to answer anything

Answering 01

One LLM key, any provider

The brain is model-agnostic: Anthropic, OpenAI, OpenRouter, or a model on your own hardware for $0.

How the routing works

OpenRouter is a gateway: one key reaches hundreds of models from every major lab. The choice is per team and lives in the database, so it is an admin-UI switch with no redeploy, and answering, reasoning and embeddings are picked independently. Without a key the dashboard still works, but the query box cannot answer.

Answering 02

Embeddings + pgvector

Load the vector schema once, then pick an embeddings model.

Why skipping this is risky

Skipping it is quiet. Search silently falls back to keyword matching only, so "what did we decide about pricing?" stops finding the thread that never says "pricing". Nothing warns you: it just answers worse.

Context engine included; optional sidecar below

Context engine

Neo4j + Graphiti, the context engine

Powers the narrative arcs on Pulse, the learning panel, and graph-grounded answers.

How extraction is configured

Graphiti uses the provider and extraction model you configure in Team Brain Admin, so it needs no separate provider key. It powers the narrative of the week as well as graph-grounded answers.

Optional sidecar

Python 3.11+ with uv

Only for the sidecar connectors: Notion, Google Drive, Confluence, RSS, web pages and local files.

What needs no Python at all

Slack, GitHub, Linear and Plane run inside the app and are configured in the admin UI. The sidecar pulls on your own infrastructure and pushes over the same API, so those credentials never reach the brain.

What it costs

Every connector API is free. You pay for embeddings and LLM calls. The context engine uses the same provider configuration as Team Brain. Self-host the model and all of it is zero.

How content gets in

Two paths, and either works on its own. Most teams use both.

  • Connectors. Slack, GitHub, Linear and Plane are configured in the admin UI and start pulling within 30 minutes. Notion, Google Drive, Confluence, RSS, web and local files run through the Python sidecar.
  • The aios CLI. Each person pushes tier-tagged content from their own workspace. Optional: the connectors feed the brain on their own.

What the installer does

In this order. Two of them are yours: choosing the answers, and connecting a source.

  1. 1Check prerequisites and validate your keysautomatic
  2. 2Generate the application and graph credentialsautomatic
  3. 3Start Team Brain, Graphiti and both databasesautomatic
  4. 4Create your team and admin loginautomatic
  5. 5Verify the result and hand you the loginautomatic
  6. 6Connect one sourceyours, ~4 min
  7. 7Configure an extraction provider in Adminyours

The Railway template performs these steps in the hosted deployment. When it is healthy, create an API key in Account and return to aios onboard; Create resumes the same validated Join flow and saves nothing until GET /api/v1/me succeeds.

Read the full setup guide Connectors and credentials

Joining a team that already runs one? You do not need any of this, you need an invite. Start at the quickstart instead.