- Reliability 6/20
- Static review sees only the README and server.json; no repository source, CI workflows, or committed tests were provided. The README claims a '15/15 endpoint smoke test' and a 'retry budget', but these are unverified assertions and, per the calibration rules, earn no credit. The manifest declares stdio transport, 18 tools, 4 prompts, 2 resources, yet tool parameters appear only as prose (e.g. palavraChave, somenteVigentes) with no JSON schema, so declared tools cannot be confirmed to match real behavior. Startup depends on npx fetching @licinexusbr/mcp pinned at 0.2.2, with no verifiable lockfile; the README itself admits PNCP may take 15-30s, and error handling is a textual promise only. Score 6: the happy path is plausible but tests and edge-case evidence are entirely absent, and nothing supports a higher score.
- Security and permissions 8/20
- No red lines observed: no malware, credential theft, covert exfiltration, unconfirmed destructive default operations; all data sources are public APIs (PNCP, BrasilAPI/MinhaReceita) and CNPJ_PROVIDER is explicitly marked non-secret and key-free. The README states clear scope boundaries (no private infrastructure, no proprietary data) and points to SECURITY.md for responsible disclosure. Deductions: aside from README self-description, there is no source-level evidence that operations are strictly read-only, that no local persistence occurs, or that network calls stay within the declared domain; execution via npx -y installs an npm package at runtime, an inherent supply-chain trust surface without committed integrity or lock verification. Score 8.
- Maintenance 10/20
- Signals: MIT license, 78 stars, 4 open issues, not archived, manifest and npm version aligned at 0.2.2, and a phased roadmap with a public release (v0.1.0). This suggests organized maintenance and clear ownership (Licinexus). Deductions: no source, commit history, release timeline, dependency-update policy, or issue-response record is verifiable; 'best effort, no SLA, triage within 7 days' is a weak commitment; the roadmap includes incomplete phases (Fase 7) and a suspicious future date (11/05/2026), reducing credibility. Score 10: active with clear license but governance and versioning evidence is thin.
- Documentation 12/20
- Documentation is layered and readable: bilingual entry points, install examples for six-plus clients (Claude Desktop, Cursor, Continue, Cline, Zed, programmatic SDK), troubleshooting for absolute-path npx, EACCES, version caching, timeouts, logging (LICINEXUS_LOG_LEVEL) and localized errors (LICINEXUS_LANG), plus tool/prompt/resource tables and a scope statement. Deductions: the 18 tools lack parameter types, required flags, defaults, pagination, and return shapes; rate limits, data freshness, and accuracy disclaimers are undisclosed; the text contains a future date and an unverified '15/15 endpoints' claim. Usable but with hidden assumptions; score 12.
- Setup experience 10/20
- The install path is concise: Node >=18 plus an MCP client, with a short JSON snippet using command npx and args -y @licinexusbr/mcp, no keys or database, covering config file locations and UI entry points for mainstream clients, plus thorough troubleshooting. Deductions: no in-repo source or executable verification, so package usability cannot be confirmed; reliance on online npx fetching is a fragile step; Claude Desktop on Linux is not officially supported; the walkthrough rests on README claims rather than checkable evidence. Well below the static ceiling of 15; score 10.