MCP

An agent connects with your personal API token, or with the acting token of a persona. Application tokens and visitor tokens are refused. The tools an agent gets are listed in the MCP reference on the help site.

What an agent can do

The tools come in groups, and each group covers one part of the hub:

• Persona storage gives a persona its own small data: key-value pairs, counters, lists, memories, events, queues and keywords.
• Discussions searches threads, lists categories, and reads your bookmarks and recent notifications.
• Profile reads what you have recorded about yourself, such as your jobs, education, locations, contacts, relationships, tags and dates, and searches your notes and data.
• Personas lists and reads personas.
• Teams lists, adds and removes team members.
• Monitored sites reaches your site groups, sites, endpoints and checks.
• Projects lists projects, iterations and work items, changes a work item's fields, and reads and writes its comments.
• Persona knowledge searches what a persona has been given to know.

Outside the groups, an agent can write a research paper, and writing the same paper again updates it rather than making a second. It can search the web. Each integration you have connected adds tools of its own, named for the provider and the action, and each outside MCP server a persona registers adds that server's tools under its name.

Reaching the rest of the API

When the token an agent brings carries a REST scope, the agent also gets a gateway to the whole API. With it an agent can search the OpenAPI document for a route and read what that route takes and returns. It can see what its own token may reach, and then make the call.

A call through the gateway is replayed through the hub's own front door, with a short-lived token good for two minutes and an origin that marks it as coming from MCP, so it passes every check a browser's call passes. A call is given 30 seconds. If it runs out, the result is reported as indeterminate, because the call may have landed.

The checks run in order. First, can an agent reach that endpoint at all; if not, 403. The reachable set is the agent allowlist, markdown content, persona memory and search, plus whatever the token's scope names. Then, does the endpoint exist; if not, 404. Then, is it in the token's scope; if not, 403. A scope of * reaches every route that can be scoped, and nothing more.

A persona acting for you

When a persona connects with an acting token, the hub first checks that the persona may act at all. Every tool call is then held to what the persona has been granted, cut down to what you yourself may do, and never below a hard floor of things no persona may do. A call that has consequences is parked until you approve it. The personas pages cover this.