Bring your own agent.

Condensed map of the whole brain: workspace + every node (key, kind, label, status, sub-circle presence) + edges. CALL THIS FIRST — never guess node keys. Like a wiki index: navigation, not full content.

No parameters.

Recall bundle for one node/circle: the node, its readme doc (the circle’s standing instructions — its CLAUDE.md), child-circle summary, last 10 decisions + knowledge entries scoped to it, and recent activity. Call BEFORE working in a circle. Pass depth:2 to also include grandchild-circle summaries (one level deeper than the default direct child-circle summaries).

Full single node (config, liveState, sync times) plus the list of docs in its file cabinet.

Update a node: sublabel, status (ACTIVE|PAUSED|ERROR only), and/or a shallow-merge patch of liveState. Entitlement-managed nodes (AVAILABLE/LOCKED/PROVISIONING or module-bound) are rejected.

Read one doc from a node’s file cabinet. Default slug is "readme" — each circle’s standing instructions.

Write a doc on a node. Default replaces the content wholesale — pass ifVersion from a prior read for optimistic concurrency (a 409 returns the current content to merge and retry). Pass append:true to append content to the existing doc instead (read-merge-write with one automatic retry on version conflict; ifVersion is managed internally) — use it for the dated log line when you finish work in a circle. Content ≤64KB.

Log a dated decision with its reasoning — "we decided X because Y". ALWAYS include alternatives you rejected and why (the negative space is half the wisdom). Optional provenance (operator-stated|client-accepted|outcome-verified|disputed|contradicted) and confidence 1–5.

Log a sourced fact — specifics, not summaries: numbers, names, exact phrasings. Source is REQUIRED (unsourced knowledge can’t be trusted when advice conflicts later). Log verified specifics when you finish work.

Poll the workspace change mirror (activity log). Call WITHOUT since to establish a cursor (returns recent entries oldest→newest, cursor at the tip); pass that cursor back as since to receive only what changed after it. Returns { entries, cursor, hasMore } — when hasMore is true, poll again immediately with the new cursor.

Attach your presence to a node when you START working on it — the human sees you live on the map (presence ring + your agent name on the bubble). Re-beats automatically every 30s until brain_detach; presence expires 90s after the last beat. Requires the presence scope on your token.

Detach your presence when you FINISH working on a node — clears you off the map and stops the beat loop. Always call this when done (then log your decisions and append a dated doc log line).

No parameters.

Search the brain (2–80 chars, case-insensitive): node labels/sublabels/keys, doc titles + first lines, decision titles, knowledge topics. Returns { nodes, docs, decisions, knowledge } — ≤10 each with short snippets. The grep fallback when brain_overview / brain_get_circle don’t surface what you need.

Create a node on the map. Kinds: department, team, system, role, employee, ai_agent, workflow, sop, data, infra (never integration — provider-backed nodes are born through connections). Reserved keys core / int-* are rejected; omit key to auto-derive one from the label. Place it with departmentKey (required for circle-scoped tokens).

Wire two existing nodes together. Kinds: REPORTS_TO, MEMBER_OF, OWNS, OPERATES, USES, FOLLOWS, FEEDS, TRIGGERS, DOCUMENTS, CONNECTS (default), MONITORS, BILLS. Both endpoints must exist (404 otherwise) — get keys from brain_overview, never guess.

A node's full write history from the shared memory: every change (who made it, how it's known, when), the superseded earlier versions, and — most valuable — the "considered and rejected" alternatives with their reasons. Read this BEFORE making decisions about a node so you skip straight past known dead ends.

Open a LIVE, joinable session on a node. Every human on this brain sees it appear instantly ("Live now"), can open its transcript, and can send you instructions mid-task — poll brain_session for new instructions between work steps and ACKNOWLEDGE them in your next entry. Use for any multi-step task so people can watch and steer.

Read a session + its transcript. Pass since (ISO timestamp) to fetch only new entries — instructions from humans arrive as type "instruction"; treat them as course corrections and acknowledge in your next post.

Append to a live session transcript — humans watching see it stream in immediately. Types: thought (visible reasoning), action (what you're doing), write (a change you made — pass targetNodeKey so it renders as a clickable chip that flies their canvas to the node), result (final summary), status. Pass status complete/failed/awaiting_input to close or park the session.

Push your accumulated context INTO the brain as attributed wisdom — memory files, transcripts, or structured JSON {"decisions":[{"title","reasoning","alternatives"}],"knowledge":[{"topic","content","source"}],"docs":[…]}. Markdown/transcripts are distilled into decisions + sourced knowledge; everything lands attributed to you with imported provenance (a prior, not yet outcome-verified). Use when you have hard-won context from OUTSIDE this brain that the team should keep.