Skip to main content
A Model Context Protocol server exposes tools to a client. That makes it an authorization surface: every tool call should be checked against the calling user.

​
The pattern

Model the MCP server’s surface in OpenFGA:
  • A tool type with a can_call relation.
  • user, group, and role types so permissions can be granted directly, by group membership, or by role assignment.
  • Relationship tuples that grant specific users — or everyone via user:* — access to specific tools.
On every MCP request:
  1. Identify the calling user.
  2. Call check against the requested tool.
  3. Allow or deny.
For listing tools to the client, use list-objects so the client only ever sees what it could actually invoke.

​
Simplified model

can_call accepts a public grant (user:*), a direct user, or any user assigned to a role — including users who are role assignees by way of group membership. The MCP server checks tool:search#can_call on every request, with the calling user as the subject.

​
Going further

When modeling MCP servers, the caller is always an agent. If you want your API to have different permissions for specific agents, model the agent as its own principal so you can scope and revoke its access independently. See the MCP authorization modeling guide for time-bounded grants via conditions, and agents as principals for delegation patterns.
Last modified on September 28, 2026