- Reliability 5/20
- The manifest is internally coherent (schema, name, version 0.3.5, npm package @jshookmcp/jshook, and matching repository/homepage), and the README describes a plausible stdio plus HTTP-daemon MCP server with profile-based tool loading. However, this review received only the README and the manifest; no source code, tool definitions, tests, or CI workflow files were provided, so the 668 declared tools cannot be confirmed to exist and the init/tool-list handshake cannot be verified. Error handling and edge cases are undocumented. Points are deducted for the absence of any verifiable execution evidence; the claim that the registry snapshot is 'checked in CI' is unverifiable, and uncertainty lowers rather than raises the score.
- Security and permissions 5/20
- No install example contains real tokens or secrets; the only auth value shown is the variable name MCP_AUTH_TOKEN. The daemon documents a loopback bind and per-client session isolation, which are positive data-boundary signals. But this is a dual-use security-analysis toolset (MITM proxy, process/memory scanning, syscall hooks, browser instrumentation, exploit-dev domain), and the supplied material shows no least-privilege defaults, no confirmation gates for destructive operations, no credential-handling policy, and no data-flow disclosure; the claimed destructiveHint annotations are unverifiable. Heavily deducted for missing permission/confirmation/scoping evidence. No clear red-line violation is evident from the material itself: no malware, credential theft, covert exfiltration, or default irreversible remote commands, and the install examples contain no real secrets.
- Maintenance 10/20
- Positive signals: the repository is not archived, has an AGPL-3.0 license, uses semver-style versioning (0.3.5), references a CI-checked generated registry snapshot, includes recent runtime notes, and points to a docs site. Deducted because the material contains no commit history, release dates, dependency-update evidence, issue-response evidence, security-response channel, or governance documentation, so the maintenance cadence cannot be verified. Star count is treated only as a discovery signal and adds no points.
- Documentation 12/20
- The README is reasonably layered: copy-paste client configs for Claude Desktop/Cursor, a Windows npx.cmd note, environment-variable explanations (MCP_TOOL_PROFILE, npm_config_omit, SEARCH_VECTOR_ENABLED, MCP_AUTH_TOKEN), daemon setup and usage, architecture notes, a generated tool/domain registry snapshot, plus links to a Chinese translation and to docs, getting-started, configuration, and tool-reference pages. Deducted because all tool parameters live behind external links not supplied for this review; there is no troubleshooting section, no cost or usage-limit disclosure despite LLM-powered features, and no runtime error semantics.
- Setup experience 12/20
- Setup is concrete and few-step: a copy-paste npx-based client config for a mainstream client, an env var to skip heavy optional native dependencies by default, a Windows workaround, and a documented HTTP-daemon alternative for URL-based clients with bind and auth notes; the Node 22.12+ requirement is explicit. Deducted because the provided material contains no real CI workflows or committed tests proving the install path works, so the score must stay well under the 15-point static ceiling; optional native packages (ONNX/Z3/Binaryen/Camoufox/Playwright) may still make installs fragile on some platforms, and the @latest pin is not reproducible.