Data Commons MCP Server

Data Commons is Google's open knowledge graph that normalises public statistics from hundreds of official sources — census bureaus, health ministries, the World Bank, the UN system and more — into one schema where every indicator shares a variable ID and every place shares an entity ID. The MCP server puts that graph behind six tools, so an agent can answer a statistical question by retrieving numbers rather than recalling them.

The tools split discovery from retrieval, which is what makes the graph tractable for a model that cannot see it all at once. search_indicators and search_child_indicators find the right statistical variable from natural language and walk down a topic hierarchy; get_variable_metadata explains what a variable actually measures, in which units, at which periodicity; then get_observations, get_child_observations and get_multi_entity_observations fetch the values — for one place, for every child of a place (every county in a state, say), or for several entities compared side by side.

The framing worth keeping in mind: this is a citation layer as much as a data layer. Asking a model for a country's under-five mortality rate gets you a plausible number; asking Data Commons gets you a number attached to a named source and a reference period.

Running it. The hosted service at https://api.datacommons.org/mcp needs no deployment — just a free Data Commons API key from apikeys.datacommons.org, sent as an X-API-Key header. For local work or an organisation's own Custom Data Commons instance, the datacommons-mcp PyPI package runs under uvx in streamable-HTTP mode (uvx datacommons-mcp serve http) or stdio mode (uvx datacommons-mcp serve stdio), with DC_API_KEY in the environment. Note that the hosted endpoint serves base datacommons.org only; a Custom Data Commons deployment must run its own server.

Apache-2.0. Version 1.4.0 shipped 3 September 2026, and the same MCP interface now fronts the UN System Data Commons.