The assistant inside the app is not privileged. It operates the product through a tool surface, and that surface is available to agents you connect yourself.
Anything the built-in assistant can do, an external agent can do. Anything you can do in the interface has a tool behind it.
Connecting a client
The remote MCP server is at https://mcp.themarginapp.com/mcp, over MCP
Streamable HTTP.
Two ways to authenticate:
- OAuth 2.1 with PKCE and dynamic client registration. Clients that support
it, including Claude, Cursor and VS Code, discover everything they need from
https://themarginapp.com/.well-known/oauth-authorization-serverand walk you through consent in a browser. - A personal access token, minted in Settings → Integrations and sent as a bearer token. Simpler for scripts and self-hosted automation.
Either way the credential is scoped to one user and one workspace, and every call is authorised against workspace membership. An agent cannot reach a workspace you are not in.
What the tools cover
166 tools across boards and columns, cards, labels, checklists, notes, habits, expenses and budgets, currency, focus sessions, whiteboards, Canva designs, attachments, templates, the family suite, sharing and connections, search, and delegation to other agents.
The server also exposes MCP resources, which are workspace summaries an agent can read for orientation, and prompts for common workflows.
Conventions worth knowing
Identifiers are UUIDs. Tools take and return JSON.
Card and column ordering uses fractional-index strings rather than integers. Use
move_card instead of writing positions directly.
Note content is markdown text. Whiteboard documents are Excalidraw scenes. Note tags are JSON arrays.
Read tools are safe to call freely. Write tools mutate real data, and anything
starting delete_ should be confirmed with the human first.
The efficient read pattern is previews first: search_notes and list_notes
return content previews, and deep reads use bulk fetches or offset paging. Ten
single fetches cost more than one bulk call, because each round re-sends the
conversation.
Agent visibility
An agent's writes appear in your app on its next sync, which is usually immediate. The exception is an open whiteboard, which adopts external changes on reopen rather than live. See Working offline.
What agents cannot reach
The Vault while locked, and anything marked Hide from memory. The boundary is the same one the built-in assistant respects, enforced in the same place, which is the reason it holds for both.
Availability
External MCP access is a paid-tier feature. The machine-readable entry points,
/llms.txt and /llms-full.txt and the discovery documents, are public.