🤖 MCP for app users

An app’s MCP endpoint is https://mcp.fleetless.dev/<app_identifier>. The person signs in through that app and approves the client. The client sees nine tools: robots, datapoints, jobs, assets, camera, commands, docs, collections_read and collections_write.

Each call names an action, such as robots.describe or commands.call_service. Call docs.describe first to get the full parameters, conditions and possible refusals for an action.

The list stays the same for every app user. The person’s current role decides what each call can do. A refusal can say forbidden, capability_required or tier_required; a listed tool is not a grant.

The role is the person’s organisation role. robots.describe lists only what it reaches, including whether it holds job_history and assets, and robots.list leaves out a robot where it reaches nothing.

collections_read and collections_write reach the records of the app’s own collections: the ones the person’s role may read, create, change or delete. Start with collections_read.definitions, and see Collections.

docs has four actions. Start with docs.search.

• describe gives the exact parameters of an action.
• search finds sections by words or identifiers and returns each hit with its page, heading path, anchor and a short snippet.
• outline lists the pages you can read, or one page’s headings with their anchors.
• read returns a page, or, with section (an anchor from search or outline), only that section. A long answer stops at a section boundary and names the next section to read.

An app session reaches only this page, the five app tool reference pages and the reference of each app tool’s actions, in all four actions. Developer, SDK and team administration pages are outside that session’s docs catalog; asking for one answers not_found, as for a page that does not exist.

An app user can withdraw a standing MCP grant in the app. A developer can withdraw that user’s grant too. Withdrawal cuts the client’s next call.

The app’s sign-in and consent screen use the pending interaction id, then send the browser to the URL returned by approval or denial.

AI agents that control robots can make mistakes. Supervise every motion and keep a way to stop the robot at once. Read the Safety notice.