The bet here is that the hard part of public statistics is interpretation, not retrieval — so the design puts margin-of-error thresholds, coverage bias and temporal validity in the path of the answer rather than in a footnote nobody reads.
An independent, open-source attempt at a statistical consultant for US Census data, built on the argument that Census has a pragmatics problem rather than a search problem: knowing which series answers a question, and how far it can be pushed, matters more than finding it. The v3 architecture is a pure Python MCP server with a pragmatic rules engine, and the repository states plainly that it is an active research rebuild.
- A pragmatic rules layer carrying fitness-for-use constraints — margin-of-error thresholds, coverage bias, temporal validity and source selection — so an answer arrives with the caveats that decide whether you can use it
- Direct Python calls to the Census Bureau APIs rather than a mirrored copy of the data
- A knowledge base of methodology documentation behind the guidance, drawn from Census methodology sources
- The quarry extraction toolkit shipped alongside, so the same method can be applied to other statistical agencies' methodology documents
Packaged as `census-mcp-server` with a `census-mcp` entry point. It is not affiliated with, endorsed by or sponsored by the U.S. Census Bureau or the Department of Commerce, and data retrieved through it stays subject to the Census API terms of service. The v3 rebuild is in progress and the architecture documentation is marked as still coming, so read the repository's current state before depending on it. The quarry toolkit is a separate exercise: it needs a Neo4j instance, an LLM API key and your own methodology PDFs.