The documentation writes itself.
Connect Third Brain to the tools your team already uses - Claude, Cursor, your agents - and every decision and answer is captured into a governed brain your whole company can search. Permissions enforced in the query. Works with every model.
Free and open source · Apache-2.0 · Bring your own model keys
We're moving on-call to a weekly rotation with a secondary. Documenting that now so the team has it.
Captured
On-call rotation change
Agent-writtenOn-call moves to a weekly rotation with a secondary…
Engineering / Decisions
Fig. 1An agent captures a decision as it works - a teammate finds it seconds later, scoped by permissions.
Works with Anthropic Claude · Google Gemini · any OpenAI-compatible model · Claude, Claude Code & Cursor via MCP
A brain that fills itself in
The knowledge your team creates with LLMs, captured into one governed source of truth - searchable by every model, the moment it's written.
The documentation writes itself.
Point your agents at Third Brain over MCP and, as they work, they call add_knowledge and update_knowledge to capture decisions, answers and notes into the right collection. The doc gets written without anyone stopping to write it - and the secret scanner gates what an agent is allowed to persist.
MCP add_knowledge / update_knowledge
Permission is enforced in the query.
Every query - a read or a write-back - is filtered by document-level access grants, server-side, before any chunk reaches the model. Immutable audit logs, per-key rate limits and usage metering keep a complete record of who asked what, and what the agents wrote.
scope = viewer+ · enforced in SQL
Any model. Zero lock-in.
Native connectors for Anthropic and Google Gemini, plus any OpenAI-compatible endpoint - OpenAI, Azure OpenAI, Ollama, vLLM or your own gateway - all behind one plain-httpx layer. Drop-in /v1/chat/completions and /v1/embeddings plus a native MCP server mean your existing tools just work, and you can swap providers without re-indexing a single document.
Anthropic · Gemini · OpenAI-compatible
Connect. Work. Documented.
The knowledge captures itself as your team works - and every read and write is scoped to what the caller is allowed to see.
Connect
One command wires the tools your team already uses into the brain: npx third-brain-mcp connect signs you in, then install claude, cursor or claude-code does the rest.
Work
Your agents search the brain for grounded answers and write decisions, answers and notes back to it as they go - no one stops to author a doc.
Documented
Review and track it all on the dashboard. An Agent-written filter shows exactly what your agents captured, and every query is scoped to what the caller may see.
Before any chunk reaches a model, the permission gate drops everything the caller can't access.
scope = viewer+ · enforced server-side
Watch it write the doc.
Three minutes against the running product: ingest a document, ask a grounded question, watch an agent capture a decision mid-chat - then watch the same question return nothing to someone who isn't allowed to see it.
Fig. 2The whole loop end to end - capture, governed retrieval, and the knowledge graph the documents form on their own.
Works with what you already run.
One command wires Claude, Claude Code, Cursor or your agents into the brain over MCP. Generate with Anthropic, Gemini or any OpenAI-compatible endpoint. No lock-in.
Then install claude, cursor or claude-code to wire in your client.
Generate with
Anthropic · Google Gemini · OpenAI · Azure OpenAI · Ollama · vLLM · Any OpenAI-compatible endpoint
Connect from
Claude · Claude Code · Cursor via MCP · Any OpenAI SDK
Capture from
Agents via add_knowledge · Files · Images · Text · URLs
POST/v1/chat/completions
POST/v1/embeddings
GET/v1/models
POST/api/v1/device-auth· CLI sign-in
MCPadd_knowledge· write-back
Slack, Drive & Notion connectors are on the roadmap.
Governed by default.
Third Brain treats access control as a first-class primitive, not an afterthought. Permissions are enforced at retrieval time, so a model only ever grounds its answers in data the caller is cleared to read - and the same gate governs what agents write back, while a secret scanner keeps credentials out of the brain.
Apache-2.0 licensed, so the enforcement path is yours to read, audit and change.
Document-level RBAC
Owner, admin, editor and viewer roles plus per-resource grants for users and teams.
Org isolation
Strict tenant scoping on every query - data never crosses organization boundaries.
Full audit trail
Every search, ingest, agent write-back and admin action is recorded with actor, IP and timestamp; per-key rate limits and usage metering included.
Scoped API keys
Least-privilege keys with read/write scopes, per-minute rate limits and act-as-user delegation; device-code sign-in mints them admin-approved.
SSO & SCIM
SAML 2.0 and OIDC sign-in with just-in-time provisioning; SCIM 2.0 provisions and deactivates users, and identity-provider groups arrive as teams.
Encrypted secrets
Connector credentials are encrypted at rest; API-key secrets are shown once, then hashed.
Self-hosted
You run Third Brain on your own infrastructure with your own provider keys, so your documents, embeddings and queries never leave your perimeter.
Give your company a brain that writes itself.
Third Brain is free and open source under Apache-2.0. Clone it, run it on your own infrastructure, and let your agents capture the work as they go - with permissions built in from the first query.