Shared Brain / Connect your agent

Setup

Connect your agent.

One MCP endpoint per workspace. Sign-in, permission approval and a test write and read happen once, and then your agent shares the workspace memory with every other agent you have authorised.

Let your agent do the setup

Copy this instruction into an MCP-capable agent. It asks for your approval before it connects, and it verifies that a write and a read both work before reporting success.

agent instruction 
Read the Shared Brain setup guide. Explain the permissions you need,
guide me through sign-in, connect to the workspace I approve, and verify a
test write and read. Do not save credentials or unrelated private information.

Nothing is saved automatically. The agent must call the memory tools, and the workspace you approve is the only one it can reach.

The manual setup flow

  1. Create or select a workspace. A workspace is the boundary an agent can read and write. Its endpoint is /mcp/w/<workspace_id>.
  2. Choose the agent and its permissions. A new agent starts with memory.read. Add memory.write only if it needs to contribute.
  3. Approve the grant. A first request is not served until a human approves the grant in the console. The path selects a workspace; it does not grant access to it.
  4. Connect the client. Point an MCP host at the workspace endpoint and complete the OAuth sign-in. Credentials belong in the host’s credential store, never in a URL or a document.
  5. Verify with a test. Run describe and its setup check. The check reports, separately, whether the request authenticated, whether the grant’s scope was verified, and whether a keyword read succeeded. When write setup was delegated, it also runs a write dry-run in a rolled-back transaction — a dry run never counts as a durable write.

A successful login is not a successful integration. Authentication, scope, the read check and the write dry-run are reported separately, and a dry run is not a durable write. The integration is only proven when a real contribution is accepted and a second authorised agent retrieves it.

Review is a state, not a suggestion. If the workspace requires review, a contributed record waits in the review inbox until a person accepts it. The contribution is not retrievable until then.

What to expect after setup

A fresh session or a context reset

The agent has no memory of the last session. It should call recall at the start of a task rather than reusing old context. Retrieved text is treated as untrusted data, never as instructions.

A correction

A correction creates a new revision and leaves the previous one visible as superseded. The next retrieval returns the current revision. Clients that subscribe to the change feed are notified and re-sync; a correction does not reach into a model’s context on its own.

An outage or a stale cache

Errors carry a code, a suggestion and a retryable flag. On resync_required, a client discards its scoped cache and re-syncs from the change feed. On unavailable, a client may retry with bounded backoff or continue with keyword-only retrieval where the deployment offers it.

Revocation

Revoking a principal ends its access going forward, and caches that follow the change feed are told to re-sync. It cannot remove content already disclosed to a third-party model, client, log or export — see the security page.

Client examples

terminal · tested 30 Sept 2026 
# Add Shared Brain as a remote MCP server
claude mcp add --transport http sbrain \
  https://sbrain.app/mcp/w/<workspace_id>

# Then sign in: run /mcp inside Claude Code

Only the clients on the tested integration matrix carry a “tested” status. The others are published setup examples, not verified integrations.