data.gouv.fr MCP Server: French Open Data in Any AI Chatbot

What the data.gouv.fr MCP Server Does and Who Uses It

data.gouv.fr is the French national open data platform, hosting datasets from government ministries, statistical agencies, and regional administrations. Browsing it manually or building queries against its API requires knowing the dataset identifiers, the available filters, and the format of the response. The MCP server removes that friction for users of AI chatbots by translating natural-language questions into data.gouv.fr API calls and returning the results in the conversation.

A user can ask questions in French such as "Quels jeux de données sont disponibles sur les prix de l'immobilier?" (What datasets are available on real estate prices?) and receive an answer drawn directly from the platform's catalog. The server was published by the official data.gouv.fr team ([email protected], per pyproject.toml) under the MIT License and reached v1.0.0 in September 2026.

The tool is most useful for journalists, researchers, policy analysts, and data engineers who routinely work with French public data and want to incorporate dataset discovery and analysis into an AI-assisted workflow.

MCP Protocol and How the Server Connects to data.gouv.fr

The Model Context Protocol is a standardised way for AI clients to call tools provided by external servers. The data.gouv.fr MCP server exposes a set of tools that the AI client can invoke: searching the dataset catalog, fetching resource metadata, and accessing dataset contents. The AI client calls these tools transparently during conversation, so the user does not need to know which API call is being made.

The server is implemented in Python using the `mcp` library (version 1.25.0 or later) and `niquests` for HTTP requests to the data.gouv.fr API. It requires Python 3.13 to 3.15. When deployed, it listens on `http://+:8000` by default, configurable through the `MCP_PORT` environment variable, and exposes the MCP endpoint at `/mcp`.

Tracking of queries uses Matomo analytics when `MATOMO_URL` and `MATOMO_SITE_ID` are configured. Sentry is supported for error monitoring via `SENTRY_DSN`. Both are optional; the server runs without them.

Connecting Claude Code, Claude Desktop, and Other Clients

The public hosted endpoint at `https://mcp.data.gouv.fr/mcp` is the simplest starting point. For Claude Code, register it with:

claude mcp add --transport http datagouv https://mcp.data.gouv.fr/mcp

For Claude Desktop on macOS or Linux, add the following to `~/Library/Application Support/Claude/claude_desktop_config.json` (macOS) or the equivalent Linux path:

{
  "mcpServers": {
    "datagouv": {
      "command": "npx",
      "args": ["mcp-remote", "https://mcp.data.gouv.fr/mcp"]
    }
  }
}

On Windows with Claude Desktop, if the server appears in the list but never connects, the README explains that Claude may be using its built-in Node.js runtime which cannot find a globally installed `mcp-remote`. Adding `"isUsingBuiltInNodeForMcp": false` at the root of the config file and restarting Claude Desktop resolves this.

For Cursor, the configuration uses `"transport": "http"` in the mcpServers object with the hosted URL. For Gemini CLI, the key is `"httpUrl"` in `~/.gemini/settings.json`. The README documents 16 client configurations in total, including VS Code, Windsurf, HuggingChat, Kiro IDE, Kiro CLI, IBM Bob, and Mistral's Le Chat.

Self-Hosting with Docker

The repository includes a Dockerfile and a docker-compose.yml for self-hosted deployments. The docker-compose configuration exposes port 8000 by default and reads environment from the host:

{
  "mcpServers": {
    "datagouv": {
      "url": "https://mcp.data.gouv.fr/mcp",
      "transport": "http"
    }
  }
}

The `.env.example` file in the repository shows the available configuration variables: `MCP_HOST`, `MCP_PORT`, `MCP_ENV`, `DATAGOUV_API_ENV` (set to `prod` for the production data.gouv.fr API, or another value for demo or staging), and `LOG_LEVEL`. Two SSRF protection variables, `URLS_ALLOW_LOCAL` and `URLS_ALLOW_PRIVATE`, default to false on the hosted server. The README notes these should only be enabled in local tooling where the server needs to fetch specifications from localhost or a private network.

The Dockerfile builds from `astral/uv:python3.14-trixie-slim`, installs git, syncs dependencies with `uv sync --frozen`, and runs `python main.py` as the entrypoint.

What the MCP Server Cannot Do

The server provides read-only access to the data.gouv.fr catalog and its published datasets. It cannot publish datasets, modify metadata, or authenticate as a data producer. Users who need to upload or update datasets must use the data.gouv.fr API directly.

The quality and freshness of the data returned depends entirely on data.gouv.fr's own catalog. If a dataset has not been updated by its publisher, the MCP server cannot compensate. Some datasets are available as file downloads rather than queryable APIs; in those cases the MCP server can locate and link to the resource but may not be able to return the actual data inline.

The public hosted instance has no stated rate limits in the README, but it is a shared service. High-volume programmatic use should be done against a self-hosted instance rather than the public endpoint.

Maintenance and Licence

The last push to the main branch was on 2026-09-08 and v1.0.0 was released on 2026-09-04. Earlier versions (v0.2.x) had been in active development since at least early 2026. The project is maintained by the official data.gouv.fr technical team at the French government's open data initiative, which provides a strong continuity signal.

The repository uses CircleCI for continuous integration; the badge links in the README point to the build status. The project is licensed under the MIT License. Test coverage is in the `tests/` directory using pytest with asyncio support. The README includes a feedback link to a Tally form for users who want to report issues or request features.

Editorial conclusion

The data.gouv.fr MCP Server is the right tool for anyone using Claude, ChatGPT, Cursor, or another MCP-compatible client who wants to query French government datasets through conversation rather than through a browser or API. The public hosted instance requires no setup. Teams that need to query the demo or staging environments of data.gouv.fr, or that need analytics over who queries what, should self-host the Docker image and configure the `DATAGOUV_API_ENV` environment variable to point at the target environment.

Frequently asked questions

Official sources

1. datagouv/datagouv-mcp on GitHub
2. License: MIT
3. Project website
4. README
5. Releases