France Opens 74,000 Public Datasets to AI Agents via Official MCP Server

March 2, 2026Provided by Utku Ege Tuluk

On February 25, 2026, France’s national open data platform data.gouv.fr launched an official Model Context Protocol (MCP) server, giving AI chatbots and agents direct, structured access to over 74,000 public datasets — no API key required. The server, developed by the government digital agency Etalab, is one of the first examples of a national government formally bridging its open data infrastructure with the emerging MCP standard.

Illustration of a bridge connecting government open data and AI neural networks — Illustration generated by AI

What Does the MCP Server Do?

The MCP server acts as a read-only intermediary between AI models and data.gouv.fr’s massive repository of public datasets. It exposes seven tools that any MCP-compatible AI client can call:

search_datasets — keyword-based dataset discovery across the entire platform
get_dataset_info — retrieve detailed metadata (title, description, organization, tags, license)
list_dataset_resources — enumerate files within a dataset
get_resource_info — access specific file metadata (format, size, MIME type)
query_resource_data — extract structured data via the Tabular API
download_and_parse_resource — handle files in unsupported formats or sizes
get_metrics — obtain visit and download statistics

Featured datasets include the Sirene business registry, deceased persons records, and property valuation requests — all queryable through natural language via any connected AI assistant.

Technical Architecture

The server is built in Python using Anthropic’s official MCP SDK, with Streamable HTTP as the sole transport protocol (STDIO and SSE are not supported). It interfaces with three internal APIs: the main data.gouv.fr API for dataset and resource metadata, the Tabular API for structured data querying, and the Metrics API for usage analytics.

A public hosted instance is available at https://mcp.data.gouv.fr/mcp with no authentication required. The entire codebase is open source under the MIT license, and local deployment is supported via Docker or the UV package manager.

The server is compatible with a wide range of AI clients: ChatGPT, Claude (Desktop and Code), Cursor, Gemini CLI, Mistral Vibe CLI, VS Code, Windsurf, IBM Bob, and AnythingLLM.

Why This Matters

This launch is significant for several reasons. First, it represents one of the earliest cases of a national government officially adopting MCP to expose public data to AI systems. While MCP has rapidly gained adoption among developer tools and commercial platforms, government adoption signals a new phase in the standard’s maturity.

Second, the read-only design is a deliberate choice. The team has stated that their long-term ambition is to eventually test write capabilities for publishing new datasets, using sovereign AI models — meaning France may eventually allow AI agents to contribute data back to the national platform, with appropriate safeguards.

Third, the administration has been transparent about limitations. Officials emphasize that language models may produce “incomplete, approximate, or erroneous” responses when querying data, and that users should treat results as starting points rather than authoritative sources. They also warn about unofficial MCP servers falsely claiming association with data.gouv.fr.

Looking Ahead

The data.gouv.fr MCP server is explicitly labeled as experimental. The team is actively inviting community testing and feedback to guide future development. The broader question it raises — how governments should expose public data to AI agents — is likely to become a recurring policy discussion as MCP adoption accelerates across the public sector.

Related Coverage

Introducing Figma’s Dev Mode MCP Server: Revolutionizing Design-to-Code Workflows — Figma’s MCP implementation for design-to-code workflows
Introducing Hugging Face’s MCP Server: Connect Your LLM to the Hub via hf.co/mcp — Hugging Face’s MCP server for model and dataset discovery