evidoria

← Back to browse

Good practice Imported

data.gouv.fr MCP Server — France's Experimental Generative-AI Gateway to Public Datasets

France · Paris · See the France profile

In February 2026, France's data.gouv.fr began exposing its dataset catalogue via a Model Context Protocol (MCP) server, letting AI assistants such as Claude answer natural-language questions over official data — a cautious, still-experimental step.

data.gouv.fr MCP Server — France's Experimental Generative-AI Gateway to Public Datasets

Details

Maturity
Pilot
Promoter
DINUM (Direction interministérielle du numérique) — data.gouv.fr
Period
since February 2026 (experimental)
Keywords
open data, generative AI, government digitalisation

Context

DINUM's data.gouv.fr team has been testing a set of generative-AI features — automated short dataset descriptions, keyword suggestions, and a Model Context Protocol (MCP) server that lets chatbots query France's public data catalogue in natural language instead of manual keyword search and download.

Objectives

The team published an explicit set of AI principles: AI is treated as an assistant rather than a decision-maker, and the platform favours state-hosted and open models (such as Albert) to protect personal data.

Activities

The MCP server launched on 25 February 2026. The team warns plainly that 'language models can produce incomplete, approximate or erroneous responses' that are 'in no way an official or reliable source' and are 'difficult to audit.'

Results

data.gouv.fr itself reports the service already experiencing performance issues under high demand, with most functionality still requiring technical setup or paid AI access.

Conclusions

The MCP server is presented internally as an experiment still 'in construction' rather than a proven, at-scale capability — an honest, cautionary case of a public data service being opened to generative AI before its reliability has been established.

Implementation
Indicative cost
Low (< €50k) — No budget figures disclosed; a small experimental feature developed in-house by the national digital agency's existing team.
Time to results
Short (< 1 year) — MCP server launched 25 February 2026; still described internally as 'in construction' rather than production-ready.
Staffing & skills
DINUM (Direction interministerielle du numerique) — data.gouv.fr team

Conditions for success

  • Access to state-hosted/open AI models (e.g., Albert) to avoid exposing personal data to third-party providers
  • A published set of AI principles framing AI as an assistant, not a decision-maker
  • Transparent public communication of the tool's current limitations

Common failure modes

  • Already reports performance issues under high demand shortly after launch
  • Most functionality still requires technical setup or paid AI access, limiting real-world usability
  • Explicit warning that outputs can be 'incomplete, approximate or erroneous' and 'difficult to audit'

Where it fits

Governance type
national digital agency (DINUM)
Scale
national
Income level
high-income

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful