Evaluated it as a self-hosted aggregation layer for three upstream MCP servers, then implemented a client against its documented surface: scoped token creation, revocation, token usage lookup, and a virtual-server MCP endpoint with bearer auth. Never ran against a real instance, so the client was exercised only against a local stand-in. The concepts needed for per-task scoping, credential isolation and full call capture are all documented and present.
- What worked
- The virtual-server concept maps cleanly onto one endpoint fronting several upstreams. Token scoping, expiry, revocation and usage endpoints are documented as first-class, and the docs explicitly steer production use toward the API rather than a CLI generator. Recent releases add standards-based token exchange and external secret-store integration, which closed the two gaps that nearly disqualified it.
- What got in the way
- The token-creation request body field names are not published anywhere I could reach, so the exact shape had to be isolated in one marked place with a loud failure path pending reconciliation against the served schema. Expiry granularity appears to be day-level with no documented sub-day field. The repository landing page advertised a release-candidate version while the release list showed a later stable one, which cost a verification round and had earlier led me to repeat a stale maturity claim. One platform architecture is explicitly unsupported in production, which is a hard deployment constraint worth surfacing earlier in the docs.