MCP servers
I give AI assistants live access to public data, with provenance on every response
Ask an AI assistant for the population of Belo Horizonte in the 2022 Census and you get a plausible number, drawn from training: maybe right, maybe from a different census, with no source. I solved it differently — these servers make the assistant query the official API on the spot.
They all speak MCP (Model Context Protocol), the open standard that connects assistants to data sources. They work with Claude, ChatGPT and any compatible client, and need no installation: point the client at the address.
What I guaranteed across the six
Provenance on every response. I made each figure come with its source, the reference period, the canonical URL that reproduces the query, the real extraction instant and the licence. It is not a decorative footer: it is what makes the number checkable, citable and auditable. When the server computes something — a variation, a distribution — the result is flagged as derived, with the convention it used.
Verified protocol conformity. I brought all six to 100% on mcpscore, with a CI ratchet that blocks regression: a change that lowers the score is not accepted.
Open source under MIT. The data is not — each source carries its own licence, and I make it travel with the number.
The servers
IBGE (Brazilian statistics)
21 tools over the IBGE APIs: geography, Census (1970–2022), economic and social indicators, SIDRA tables, cartographic meshes, given names, economic activity codes and health. Ranks all 5,570 municipalities in a single call, with the full distribution instead of a top 10 that hides the skew.
Central Bank of Brazil
15 tools over three BCB sources: the SGS time series, the Focus market-expectations survey and the official PTAX exchange rates. It handles the API’s limits on its own — a 15-year daily window is sliced and merged — and compounds rather than sums the series that already are rates of change.
Brazilian Federal Senate
67 tools over the Senate’s open data: senators, bills, votes, committees, floor proceedings, the e-Cidadania participation layer and the administrative tier (expenses, procurement, staff, budget). An independent, read-only service, not affiliated with the Senate.
Medical terminologies
31 tools over ICD-11, LOINC, RxNorm, MeSH, ATC and the Brazilian CID-10 (DataSUS V2008, bundled — no network call), plus WHO’s official ICD-10 → ICD-11 transition tables. Portuguese labels come from WHO and NLM themselves; never from machine translation.
ILOSTAT (ILO)
4 tools over the International Labour Organization’s statistical database: more than 1,200 indicator dataflows on labour, by country, year, sex and age group. You search the dataflows before pulling data, so you never have to guess an indicator code.
UNESCO UIS
3 tools over the UNESCO Institute for Statistics: around 5,000 indicators on education, science and R&D, culture and communication, by country, region and year. The data release is pinned on every query, so the same question gives the same answer tomorrow.
How to use them
In any MCP client with HTTP support, point it at the server address followed by /mcp. For the IBGE one:
https://ibge.sidneybissoli.com/mcpNo installation, no sign-up, no API key. For local use, the five published to npm run with npx — for example, npx -y ibge-br-mcp.
Then just ask:
“What was the population of Belo Horizonte in the 2022 Census?”
“What does the market expect for Brazilian inflation in 2027?”
“How did each senator vote on the last recorded floor vote?”