data.gouv.fr uses AI to simplify access to public data

Summary

France's open data portal, data.gouv.fr, is piloting features powered by large language models (LLMs). These include automatically generated descriptions, keyword suggestions, and an MCP server that connects the catalog to third-party chatbots. The portal positions AI not as an autonomous agent, but as an assistant operating under human supervision.

Why it matters

data.gouv.fr is one of the first national open data infrastructures to open its catalog via the Model Context Protocol (MCP). This standard could allow any LLM-based tool to query government datasets locally. The use of government-hosted open-source models adapted by DINUM, along with a policy of not exporting data, establishes a reference point for public sector AI governance that other governments can emulate.

Highlights

• The MCP server enables Claude, ChatGPT, or Mistral to search and analyze data.gouv.fr datasets in natural language without writing custom API code.
• The skills repository provides structured LLM documentation, allowing models to locally invoke the platform's catalog, metrics, and tabular data APIs.
• All operations run on government-hosted models. No personal data is collected or sent to external services, and feedback is anonymized.

Critical take

The page reads more like a manifesto of principles than a technical explanation. Every feature is still marked as under construction. Error rates, latency times, or real-world usage metrics are not shared. The promise of ecological simplicity sits awkwardly alongside the decision to run LLM inference at a national scale. Until the MCP server and skill set are stress-tested beyond the team's own workflows, their practical value beyond the novelty of the protocol remains unproven.