- Reliability 11/20
- The manifest and README describe a stdio transport, uvx runtime and 44 tools whose names and descriptions are internally consistent; offline checks (CPF/CNPJ/NF-e validation, SPED parsing) are plausible and reproducible in principle. However, this review only has server.json and the README: no source files, tests or CI workflow definitions are supplied, so the claimed 85% coverage and CI badge are unverifiable self-assertions and do not count as execution evidence. A positive: certificate-dependent tools document their failure semantics (FiscalConfigurationError / HTTP 503) separately from transient SEFAZ outages, indicating controlled, diagnosable errors. Per the static calibration cap of 12 without verifiable tests/CI, I score 11.
- Security and permissions 14/20
- No red lines triggered: no malware, no credential theft, no covert exfiltration. The A1 certificate and password are user-supplied locally; the README states mTLS and XMLDSig are performed locally, secrets are never sent to external servers, and the password must come from a secrets manager rather than a versioned .env. The unauthenticated certificate-status endpoint returns only validity, not identity. Operations are mostly read-only consultations with no irreversible delete/pay/deploy defaults. Deductions: outbound calls (BrasilAPI, ReceitaWS, SEFAZ) are disclosed only in prose with no visible least-privilege or confirmation implementation; signing/download tools have no described confirmation step; the data-flow disclosure rests on a single untrusted document. Score 14.
- Maintenance 14/20
- Clear MIT license, consistent versioning between manifest and README (0.5.1), a CHANGELOG, a Roadmap with completed and planned releases, Discussions, a newsletter, and mentions of release automation; 295 stars and 12 open issues suggest activity. Deductions: no source is visible, so commit cadence, issue response, dependency updates and an actual security-response channel cannot be verified; the v1.0.0 maintenance-SLA item remains unfinished. Score 14.
- Documentation 13/20
- Reasonably layered docs: quick start, per-client config examples (Claude Desktop/Code, Cursor, VS Code+Continue, Docker), an environment-variable table, tools grouped by category (functional/guidance/certificate/experimental) with stated limits (e.g. no public name-based CNPJ search), certificate configuration, and SDK/CLI/REST usage. Deductions: per-tool parameter signatures and return schemas are not documented; cost/rate-limit boundaries are vague; demo transcripts are illustrative, not reviewable evidence; the competitor comparison table is unverifiable marketing copy. Score 13.
- Setup experience 15/20
- The install path is short and clear: a single `uvx mcp-fiscal-brasil` command, plus uv installation notes and an @latest refresh hint, with config snippets for mainstream clients and a Docker run command, and no API key required. That is a clean, few-step setup that should work in mainstream clients. Deductions: no verifiable in-file execution evidence (no real CI or test files visible), so per the calibration cap setup cannot exceed 15; platform/server-host compatibility nuances are not fully covered. Score 15 (at cap).