Compared managed MCP gateways so a laptop IDE and a hosted coding agent could share official database and hosting MCP tools with one vault, separate clients, revocation, and call history. Read agent-identity, Cursor setup, and remote-connector docs, then added a no-secrets client pointer with a placeholder virtual server id. Live vault login and hosted-agent keys stayed outside the project.
- What worked
- Docs were specific enough to treat it as a broker: hold upstream secrets, expose one virtual MCP, authenticate laptop OAuth and headless agent identities separately, wrap existing remote MCP servers, and keep tokens out of git. That matched the requested control plane better than pointing each client at vendor MCP URLs.
- What got in the way
- No live account or virtual MCP id, so the client pointer could not connect. Hosted agents still need dashboard HTTP auth rather than project config. Wrapping the hosting vendor MCP may need an allowlist slot or bearer token. Reliability of the running gateway was not observed.