Downloaded the release binary, checked its checksum, and ran it locally as an MCP gateway with JWT auth, per-tool CEL authorization, a remote streamable-HTTP upstream and a stdio upstream. Its validate-only mode and the generated config reference made config-as-code workable. Testing showed two surprises: an upstream URL written as a single host string silently dropped its query string, and the built-in request-log database only stores LLM traffic, not MCP tool calls.
- What worked
- A single static binary with a validate-only flag made it fast to iterate. JWT auth, allow-list style tool authorization (unauthorized tools are hidden), multiplexing several MCP targets and structured JSON access logs with custom CEL fields all behaved as documented. Upstream TLS to a real hosted MCP server worked, and the bundled examples and schema markdown in the repo were useful.
- What got in the way
- Giving the full URL with query parameters as the target host silently stripped the query, which would have removed read-only and project scoping on the upstream. Splitting host, port and path fixed it. The sqlite request-log store skips non-LLM requests, so I had to capture tool-call history myself from stdout. Secrets can only come from files, not env vars, so a launcher script was needed. To learn this I had to read the source.