Skip to main content

Design: MCP gateway and registry

Status: built, off until an operator registers a server.

Agent teams put a gateway between agents and MCP tool servers to get one place for who may call what, how often, and a record of it. Merion already had the hard parts (the egress broker with host rules, MCP body rules, DLP, taint, approvals, the vault). This adds the registry and the relay on top, so none of those are re-implemented.

As built

  • Registry (operator only). GET /v1/mcp/servers, GET/PUT/DELETE /v1/mcp/servers/{name} on the public router. An entry has url (https, or http to a loopback address; no credentials in the URL), an optional credential (a vault name the broker attaches; the agent never sees the value), tools (the tool names a tools/call may use; empty means none) and max_calls_per_minute. The name comes from the path. Stored in mcp_servers.json; set and remove are journaled (merion.mcp_server.set, .removed).
  • Relay (the cell). POST /v1/mcp/{server} on the broker port, with the same session id and egress capability headers as /v1/egress. Body: one JSON-RPC 2.0 request. Relayed methods: initialize, notifications/initialized, ping, tools/list, tools/call. Batches and every other method (sampling, resources, prompts, completion) are refused with 403. A tools/call must name a listed tool.
  • Rate limit per session and server per minute, in memory (a restart resets it).
  • Journal. Each call writes an mcp.call row with the method, tool name and outcome. Arguments and results are never journaled.
  • Then the ordinary broker. The relay builds an egress request to the server's URL and runs it through the same path as any brokered call, so the agent's own host rules, MCP body rules, DLP, taint, write limit and approvals still apply. A registry entry is necessary, not sufficient: the agent's policy must also allow the server's host, and an ask rule there holds a call for a person.

Not built

  • No per-tool "ask" in the registry itself; approvals come from the agent's egress rules for the host.
  • No discovery or health checking of servers, no SSE or streaming responses, no session ids for stateful servers.
  • The rate limit is not shared across runtime replicas.
  • Tested with unit tests and a local stub server that answers every request with ok; not against a real MCP server.