France Builds an Official MCP Server for Its National Open Data Platform - 74,000 Datasets Now Queryable by AI Agents
The French government's digital agency Etalab shipped an open-source MCP server for data.gouv.fr, giving AI agents structured access to 74,000 public datasets without an API key.
6 min read
The French government just shipped an MCP server for its national open data platform. It's open-source, MIT-licensed, requires no API key, and gives AI agents structured access to 74,000+ public datasets on data.gouv.fr - the platform ranked #1 globally in the OECD Open, Useful, Reusable Government Data Index.
This is, as far as I can find, the first time a national government has built and shipped an official MCP server for its open data infrastructure. Not a third-party wrapper. Not a community project. An official tool built by Etalab, the French government's open data division within the Interministerial Digital Directorate (DINUM), launched at a public endpoint that anyone can connect to.
TL;DR
- France's digital agency Etalab built an open-source MCP server for data.gouv.fr - 74,000+ public datasets accessible by AI agents
- No API key required, public endpoint at
https://mcp.data.gouv.fr/mcp, MIT license, source on GitHub - 10 tools covering dataset search, metadata retrieval, resource downloads, geographic filtering, schema validation, and API querying
- Experimental status - Etalab is explicit that this is a "first exploration" and wants community feedback
What It Does
The Stack
The server is built in Python using the MCP SDK (mcp[cli]>=1.8.0) with Server-Sent Events (SSE) transport. It wraps data.gouv.fr's existing REST API and exposes it through 10 structured tools that an AI agent can call without writing custom API integration code.
# Connect to the public endpoint - no API key, no auth
# Add to your MCP client config:
{
"mcpServers": {
"datagouv": {
"url": "https://mcp.data.gouv.fr/mcp"
}
}
}
Or run it locally:
git clone https://github.com/datagouv/datagouv-mcp.git
cd datagouv-mcp
uv run datagouv-mcp
Dependencies are minimal: httpx for HTTP requests, mcp[cli] for the protocol layer. That's it.
The 10 Tools
| Tool | Function |
|---|---|
search_datasets | Full-text search across all 74,000+ datasets with pagination |
get_dataset | Retrieve full metadata for a specific dataset by ID |
get_resource | Get metadata for a specific resource (file) within a dataset |
get_resource_data | Download actual data from a resource |
search_datasets_by_geographic_area | Filter datasets by geographic zone (commune, department, region) |
get_allowed_topic_values | List available topic categories for filtering |
get_allowed_schema_values | List available data schemas |
search_datasets_by_schema | Find datasets conforming to a specific schema |
get_api_resource_data | Query API-based resources directly |
get_api_resource_available_fields | List queryable fields for API resources |
The geographic tools are especially useful for French public data. data.gouv.fr organizes datasets by administrative geography - commune, department, region - and the MCP tools expose that structure directly. An agent can ask for "all environmental datasets in Ile-de-France" without needing to know the underlying API format.
What's Actually on data.gouv.fr
This isn't a token gesture. data.gouv.fr is one of the largest national open data platforms in the world:
- 74,000+ datasets from government ministries, agencies, and public institutions
- 45,000+ contributing organizations
- Covers everything: transportation schedules, healthcare statistics, environmental monitoring, education data, election results, real estate transactions, business registries, geographic boundaries
- The OECD ranks France #1 globally in its Open, Useful, Reusable Government Data Index for 2023
- All data is freely available under the French Licence Ouverte (Open License 2.0), compatible with CC-BY
Some of the most-used datasets: the national address database (Base Adresse Nationale), SIRENE business registry, cadastral data, real-time transportation feeds, COVID tracking data, and air quality measurements.
How It Compares
The MCP ecosystem is mostly private companies and community projects. Government-built MCP servers are rare.
| MCP Server | Builder | Data | Auth | License |
|---|---|---|---|---|
| datagouv-mcp | French government (Etalab) | 74,000+ public datasets | None required | MIT |
| GovInfo MCP | US community project | US government publications | API key | MIT |
| Various private | Companies | Proprietary data | API key/OAuth | Mixed |
The closest comparison is the GovInfo MCP server for US government publications, but that's a community-built wrapper, not an official government project. The French server is maintained by the same team that maintains the platform itself - Etalab's developers know the API intimately because they built it.
The no-auth-required public endpoint is a deliberate design choice. data.gouv.fr's data is already public. Adding API key requirements would just create friction without improving security. An agent can connect and start querying immediately.
Where It Falls Short
It's experimental. Etalab labels this explicitly as an "experimentation" and a "first exploration." The blog post announcing it uses careful language about this being a prototype to gather feedback, not a production commitment. The repository has 46 stars and 3 contributors as of publication.
10 tools cover the basics, not the full API. data.gouv.fr's REST API has significantly more endpoints than what the MCP server exposes. Community resources, discussions, reuses, spatial queries, and harvesting endpoints aren't covered. If your use case needs those, you're still writing custom code.
No write operations. This is read-only. You can't create datasets, update metadata, or upload resources through the MCP server. For organizations that manage data on the platform, this is a viewer, not a management tool.
French government data has French-specific quirks. Administrative geography codes (INSEE codes), schema conventions, and metadata fields follow French standards that non-French developers may find unfamiliar. The data itself is often in French. An AI agent can handle the translation, but the underlying structure assumes familiarity with French institutional organization.
MCP security remains an open question. The server is read-only and doesn't require authentication, which limits the attack surface. But the broader MCP ecosystem's track record with security isn't inspiring - one in three MCP implementations has command injection vulnerabilities according to Endor Labs. The datagouv server's Python code uses httpx for HTTP requests rather than shelling out to commands, which avoids the worst class of MCP vulnerabilities. But any tool that gives an AI agent access to external data introduces prompt injection risks.
No caching or rate limiting documentation. The README doesn't mention rate limits on the public endpoint or whether results are cached. For production use, this matters. Hammering data.gouv.fr's API through the MCP server at scale without understanding the limits is a recipe for getting blocked.
A government building official infrastructure for AI agents to access public data is a signal worth paying attention to - regardless of the "experimental" label. France is the first country to do this at the national level, and they did it the right way: open source, no authentication barrier, MIT license, public endpoint. The technical implementation is straightforward Python wrapping a REST API - nothing revolutionary. But the political decision to build it, ship it, and put it on GitHub under a government organization is the story. If this works, expect other OECD countries to follow. If it doesn't, it'll be because nobody used it - and that's a data problem, not a code problem.
Sources:
