Scored across 23 tools
Most tools target clearly distinct resources (localities, census, news, calendar, CNAE, countries), and descriptions include explicit 'Use a different tool when' guidance that resolves overlaps. However there is real conceptual overlap among the SIDRA wrappers (ibge_censo, ibge_indicadores, ibge_cidades, ibge_comparar, ibge_datasaude) and the low-level ibge_sidra, so an agent could still misselect on an ambiguous query.
Nearly all tools use a consistent ibge_ snake_case prefix with Portuguese resource/action names (ibge_municipios, ibge_estados, ibge_comparar, ibge_geocodigo). The only deviations are 'search' and 'fetch', which are deliberately named to satisfy the OpenAI Deep Research contract, a minor and justified exception.
23 tools is on the heavy side, but the domain is extremely broad (geography, census, time series, health, news, calendar, CNAE, countries, meshes, names) and a layered low-level-plus-wrapper design is intentional. Each tool maps to a distinct resource or workflow, so the count is justified though bordering on large.
The surface covers an impressively wide range: locality lookup, geocoding, administrative and thematic meshes, census/indicator/health wrappers, the raw SIDRA engine with table discovery and metadata, news, calendar, CNAE, names, countries, and comparison. Lifecycle operations (find table -> metadata -> query, name->code->details) are all present with no obvious dead ends.