Reference
MCP
The MCP endpoint: its five tools, how every read is scoped to the person asking, and citations.
Source: docs/mcp.md on GitHub, at commit 7163b59 (2026-10-16).
One endpoint, four tools, and every read scoped to the person asking.
https://mcp.livingbrain.wiki/v1/pages/mcp — Streamable HTTP, POST only, no trailing slash.
Connect
claude mcp add --transport http livingbrain https://mcp.livingbrain.wiki/v1/pages/mcp
codex mcp add livingbrain --url https://mcp.livingbrain.wiki/v1/pages/mcp
cursor mcp add livingbrain --url https://mcp.livingbrain.wiki/v1/pages/mcp
opencode mcp add livingbrain --url https://mcp.livingbrain.wiki/v1/pages/mcpThen ask it something about a repository. claude calls brain_context_for and answers with a cited brief.
Tools
| tool | what it does |
|---|---|
brain_search(query, limit?) | pages you may read holding every word of the query |
brain_page(slug) | one page's Markdown, by slug |
brain_context_for(repo, task) | a brief: the pages that mention the task, full text, cited |
brain_note(text, title?, facts?) | writes to your own private memory, redacted first; facts are temporal facts the note states (format) |
brain_facts(subject?, attribute?, as_of?) | what is true about an entity now, or as of a date, with the history of earlier values (format) |
Every page a tool quotes comes back as a numbered citation — title, scope/slug ref and URL — and initialize tells the model to cite the [n] markers.
What an agent can see
Whatever its user's scopes_for(user, Location::Dm, …) grant — the same rule Slack gets — and nothing else. No tool takes a scope argument. A page you were not granted reads as not found, not as a refusal, because "you cannot see this" is itself an answer about somebody else's page. A page scope is {shared|channel|user}-{hash of the workspace and the scope}: a page has no tenant column, so two workspaces' pages are kept apart by the scope string.
Logging in
Pass a personal access token from Settings → API tokens:
claude mcp add --transport http livingbrain https://mcp.livingbrain.wiki/v1/pages/mcp \
--header "Authorization: Bearer <token>"Or run livingbrain login, which uses the device authorization grant (RFC 8628): it prints a short code, and you approve it in a browser where the session cookie you are already signed in with is the proof. The token is returned once; the API stores only a SHA-256 hash, and a page outside the token's scope subset reads as not found.
A client that reads the metadata finds both halves itself: /.well-known/oauth-protected-resource names this endpoint and the authorization server, and /.well-known/oauth-authorization-server names the device and token endpoints. The 401's RFC 9728 challenge points at the first.
Not implemented yet: the browser authorization-code + PKCE flow, with dynamic client registration, that claude mcp add uses for zero-config OAuth. The metadata advertises the device_code grant only, so a client that only speaks the redirect flow will not complete a login on its own.