- Reliability 11/20
- README declares 11 tools, multiple image variants, and layered docs; tool catalog is internally consistent, but the supplied material contains no committed tests, CI workflows, or executable evidence verifying the init handshake and tool behavior. Cannot award above 12 under static calibration; deducted for missing verifiable test/CI evidence.
- Security and permissions 14/20
- Read-only dependency-intelligence server whose side effects are outbound calls to Maven Central and optional Context7; CONTEXT7_API_KEY is optional and passed via env, no hardcoded tokens, no delete/pay/deploy defaults, no red-line risks observed. Deducted for incomplete disclosure of outbound data-flow boundaries (Context7 doc queries) and confirmation/scoping details.
- Maintenance 14/20
- Version 3.2.2, explicit MIT license, NOTICE attribution, CHANGELOG, and documented weekly dogfooding workflow suggest sustained care; 0 open issues. Deducted because actual release cadence, commit activity, and a security-response channel cannot be verified from static material alone.
- Documentation 16/20
- Well-layered docs (setup/tools/examples/troubleshooting/architecture/dogfooding) with client config examples, disclosed limitations (not fully offline), parameter references, and license info. Deducted because linked docs' contents are unverifiable in static review and cost/quota details are only partially disclosed.
- Setup experience 14/20
- Copy-paste JSON configs for Claude Desktop and VS Code, multi-arch Docker images, clear variant guidance, and a short quick-start path. Deducted because files lack verifiable CI/install validation evidence per calibration rules, and variant/network choices add mild fragility.