Skip to main content
The Hexgate MCP server lets an AI client such as Claude Code work on a project’s policy: read the agents and the audit log, write a policy, check it, dry-run calls against it, and save it. The client logs in as you, for the projects you choose and the scopes it asks for, and you can end the login at any time.

Connect Claude Code

The server is at /mcp on the platform’s address:
  1. Register it once:
  2. Start a new Claude Code session. A server added during a session is not picked up by it.
  3. Run /mcp, choose hexgate and Authenticate. Your browser opens the Hexgate dashboard, which asks you to sign in if you are not already, then shows a consent page.
  4. Review the consent page (below) and Approve. Claude Code now lists the Hexgate tools.
There is no API key to copy: the login is a token Claude Code renews by itself.

What you approve

The consent page names the client, the scopes it asks for, and the projects you can let it reach. You tick the projects: a client only ever reaches the ones ticked, even in an organisation where you are an admin of more. Deny ends the login with nothing granted. Scopes limit what you let the client do; your own rights still apply on top. Saving a policy file needs admin rights on the project, for example, whatever the scope says. A tool called without its scope is refused and says which scope the connection lacks. To change what a client may do, revoke it and connect again.

Tools

What an agent’s code, the integrator or an end user’s request supplied (a tool description, a call’s arguments, the user id and roles a request asserts, a denial reason) comes back in fields named untrusted_*. The server marks it so the client reads it as data and not as instructions. The policy reference tool serves only the Policy pages; a link in them to any other page comes as its https://docs.hexgate.ai address. On your own machine, a server started with HEXGATE_MCP_LOCAL_POLICY_DIR also has policy_write_local, which writes policy files into that folder instead of the platform. It is never on in a deployed server; see environment variables.

Connected apps

Every login you have approved is listed under Connected apps in the dashboard’s sidebar, beside Account: the client’s name, the scopes and projects it was given, when it connected and when it last renewed its access (not on every call it makes). Check it after any consent you are unsure about. Revoke logs that client out: it can no longer renew its access and has to ask you again. A token it already holds keeps working until it expires, within 10 minutes. A login also ends on its own after 30 days without use, and in any case 90 days after you approved it.

Run it locally

To try the whole flow on your machine, the repo’s run-platform skill starts the API, the dashboard and the MCP server and registers the local one with Claude Code. The server listens on http://localhost:8080/mcp there.