Agents & API

What you see

1. API tokensCreate a token, see your active tokens (name, prefix, access, created, last used) and revoke them.
2. What it may doPropose only reads, proposes and runs AI checks; nothing it does becomes canon. Full access: edits as you can also make editorial edits (documents, pages, notes, the Library). Neither decides canon. Fixed when the token is made.
3. Connect your agentReady-made setup for Muse Code (Meta), Claude Code, Cursor and Other client, plus a small script for your own code on the Meta Model API. Your new token is filled in.
4. What your agent can doEvery tool, grouped: read canon, read sources, check, read the rest of the project, propose and run checks, edit.
5. Try askingExample requests to paste into your agent.

Create a token

1. Type a Token name that says which agent or machine uses it.
2. Under What it may do, choose Propose only or Full access: edits as you.
3. Click Create token.
4. Click Copy token now: it starts with sb_, is shown once and is stored only as a hash.

Connect your agent

1. Under Connect your agent, pick your client’s tab.
2. Click Copy and paste the config or command where the tab says.
3. The agent reaches the project at its /mcp address (Streamable HTTP), with the token in the Authorization: Bearer header.

Revoke a token

1. Click Revoke on the token.
2. Click Revoke again to confirm, or Keep to cancel; it stops working at once.

Propose only vs full access

• Both levels read everything in the project: canon, sources and their history, entities, conflicts, changes, tasks, notes, research notes, conversations, the audit trail and metrics.
• Propose only adds proposals and AI runs, and nothing it does becomes canon: add a candidate to Review, propose a change and re-run its impact, promote a research note to a candidate, start and post in Ask conversations, ask the AI critic, re-extract a document, scan for conflicts, run a continuity check or research, and measure the gold set.
• A proposed candidate must quote an imported document word for word; it waits in Review marked "Agent proposal".
• Full access adds the editorial work your role allows, made as you and shown "via agent" in the audit trail: import and write documents, drafts, notes, tasks, entity pages (not merges), library folders and tags, images, project settings and entity types, share links, and renaming conversations.
• No level decides canon. Agents read canon, check facts and propose; a person decides in CanonTracker: accepting, editing or rejecting candidates, adding a fact straight to canon, deciding and reverting changes, judging conflicts and their severity, re-checking citations and merging entities. An agent that asks for one is refused.
• If canon has nothing, the agent is told "unknown"; it never gets invented sources.

Limits

• A token belongs to you, works for one project (the one open when you made it) and always has your current role.
• Your role still applies: a viewer’s token reads and posts in Ask conversations (as viewers can in the app), but can’t propose or edit, whatever its level.
• Each token can make 120 tool calls a minute; AI tools also count toward your own hourly limits.
• Clients that support MCP prompts get two ready-made tasks: Brief me on an entity and Continuity-check a scene.
• Agent runs show in Agent activity with the model id, like any AI run.

What agents can’t do

Use the app for these: every canon decision (Review, Conflicts, Changes, re-checks, merges), your account, studios, team changes (roles, removing members, invites), API tokens, accepting invites, loading a sample (it replaces the project), audio uploads and dictation, and file downloads (portraits, images, audio, EPUB).

Who can do what

Everyone

Create and revoke their own tokens. A token can do what its owner’s role allows, within its access level.

Viewer

Tokens read the project and post in Ask conversations, like a viewer in the app.

Editor and up

Propose only tokens read, propose and run AI checks; full-access tokens also make editorial edits, as you. Canon decisions stay with people.

Roles are per studio. Roles and permissions has the full list.