One MCP server, 70 Brazilian government APIs, zero API keys

What it does

mcp-brasil is a Python MCP server that exposes 70 Brazilian public data sources — Banco Central, Câmara, Senado, TSE, DataJud, IBGE, ANVISA, a dozen state audit courts and more — as 533 tools, 131 resources and 102 prompts. You point Claude, GPT or Copilot at it and ask questions in plain language: biggest federal contracts of 2024, Selic vs. IPCA trends, a candidate’s top donors. 66 of the APIs need no key at all; the other 4 use free registrations.

The interesting bit

Two design choices stand out. planejar_consulta builds execution plans that cross-reference multiple APIs — a deputy’s expenses joined with their votes and bills — which is exactly the kind of multi-source stitching that’s tedious by hand. And the big-dataset features (TSE’s ~4M candidate records, 813k federal properties, ENEM microdata) download into an embedded DuckDB so subsequent queries run in milliseconds instead of hammering a government endpoint.

Key highlights

• BM25-based tool discovery filters 533 tools down to what’s relevant to the current context
• executar_lote fires parallel queries in a single call
• Adding a feature is just creating a folder — auto-registry, no manual wiring
• Async throughout: httpx, Pydantic v2, rate limiting with backoff
• Local datasets are opt-in via env, so the default install stays lean

Caveats

• Independent project — not affiliated with the government; each data source carries its own license, and the README explicitly tells you to read SOURCES.md and the acceptable-use policy before commercial or journalistic use
• The opt-in datasets are heavy: roughly 100MB–1.6GB downloads, with first ingest taking minutes

Verdict

If you build agents that touch Brazilian civic, fiscal or health data, this is a serious head start — the cross-referencing and DuckDB caching alone justify a look. If Brazil isn’t your data jurisdiction, move along.