- Reliability 9/20
- The README describes 14 tools, error handling, path validation, and a test command, making the happy path plausible. However, the manifest is not cached and no source files, CI workflows, or committed tests are provided, so MCP initialization/tool-list behavior, implementation parity, and edge-case failures cannot be verified. The README also claims “zero dependencies” while requiring npm-installed dependencies, and use of latest reduces controllability; points are deducted for these uncertainties.
- Security and permissions 13/20
- The README explicitly describes stdio transport, vault-root confinement, traversal and symlink checks, extension/directory filtering, and confirmation parameters for deletion and file moves, showing meaningful boundary and side-effect controls. However, there is no authentication or caller-authorization mechanism; a connected AI can read and modify all permitted content. The controls are not source-verifiable here, and evidence for network, OS-permission, and sensitive-data handling is limited, so this is well below full marks.
- Maintenance 10/20
- The repository is not archived, the MIT license is stated, and the material identifies an owning organization, a changelog link, and 13 open issues. But there is no verifiable evidence of release or commit cadence, CI, dependency-update practice, issue response, or a security-reporting channel, so sustained maintenance cannot be awarded.
- Documentation 15/20
- The README covers runtime prerequisites, installation, several mainstream client configurations, example parameters and responses for the 14 APIs, limits, troubleshooting, security considerations, and testing commands. It is substantially usable, but there is no source or manifest cross-check; authentication, the permission model, dependency/version pinning, cost, and some platform claims lack independent evidence. The “zero dependencies” wording is also inconsistent with the installation instructions, so points are deducted.
- Setup experience 13/20
- The path from installing Node.js to launching through npx, testing with Inspector, and configuring Claude and other clients is clear, short, and illustrated for multiple platforms. It still depends on npx/network retrieval of latest, client capability or subscription constraints, and local filesystem permissions, and there is no real CI/test evidence validating the connection flow. Static calibration therefore limits the score and uncertainty reduces it further.