# data.gouv.fr MCP Server

<img width="1200" height="675" alt="image" src="https://github.com/user-attachments/assets/5d20e992-349a-4b3b-9a0a-ebe308735cc9" />

> [!TIP]
> Got feedback? [Tell us about it here](https://tally.so/r/KYMboX)

[![CircleCI](https://circleci.com/gh/datagouv/datagouv-mcp.svg?style=svg)](https://circleci.com/gh/datagouv/datagouv-mcp)
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)

Model Context Protocol (MCP) server that allows AI chatbots (Claude, ChatGPT, Gemini, etc.) to search, explore, and analyze datasets from [data.gouv.fr](https://www.data.gouv.fr), the French national Open Data platform, directly through conversation.

Instead of manually browsing the website, you can simply ask questions like "Quels jeux de données sont disponibles sur les prix de l'immobilier ?" or "Montre-moi les dernières données de population pour Paris" and get instant answers.

> [!TIP]
> Use it now: A public instance is available for everyone at https://mcp.data.gouv.fr/mcp with no access restrictions.
> To connect your favorite chatbot, simply follow [the connection instructions below](#-connect-your-chatbot-to-the-mcp-server).

## 🌐 Connect your chatbot to the MCP server

Use the hosted endpoint `https://mcp.data.gouv.fr/mcp` (recommended). If you self-host, swap in your own URL.

The MCP server configuration depends on your client. Use the appropriate configuration format for your client:

[AnythingLLM](#anythingllm) | [ChatGPT](#chatgpt) | [Claude Code](#claude-code) | [Claude Desktop](#claude-desktop) | [Cursor](#cursor) | [Gemini CLI](#gemini-cli) | [HuggingChat](#huggingchat) | [IBM Bob](#ibm-bob) | [Kiro CLI](#kiro-cli) | [Kiro IDE](#kiro-ide) | [Le Chat (Mistral)](#le-chat-mistral) | [Mistral Vibe](#mistral-vibe-cli) | [OpenCode](#opencode) | [VS Code](#vs-code) | [Windsurf](#windsurf)

### AnythingLLM

1. Locate the `anythingllm_mcp_servers.json` file in your AnythingLLM storage plugins directory:
   - **Linux**: `~/.config/anythingllm-desktop/storage/plugins/anythingllm_mcp_servers.json`
   - **MacOS**: `~/Library/Application Support/anythingllm-desktop/storage/plugins/anythingllm_mcp_servers.json`
   - **Windows**: `C:\Users\<username>\AppData\Roaming\anythingllm-desktop\storage\plugins\anythingllm_mcp_servers.json`

2. Add the following configuration:

```json
{
  "mcpServers": {
    "datagouv": {
      "type": "streamable",
      "url": "https://mcp.data.gouv.fr/mcp"
    }
  }
}
```

For more details, see the [AnythingLLM MCP documentation](https://docs.anythingllm.com/mcp-compatibility/overview).

### ChatGPT

*Available for paid plans only (Plus, Pro, Team, and Enterprise).*

1. **Access Settings**: Open ChatGPT in your browser, go to `Settings`, then `Apps and connectors`.
2. **Enable Dev Mode**: Open `Advanced settings` and enable **Developer mode**.
3. **Add Connector**: Return to `Settings` > `Connectors` > `Browse connectors` and click **Add a new connector**.
4. **Configure the connector**: Set the URL to `https://mcp.data.gouv.fr/mcp` and save to activate the tools.

### Claude Code

Use the `claude mcp` command to add the MCP server:

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

### Claude Desktop

Add the following to your Claude Desktop configuration file (typically `~/.config/Claude/claude_desktop_config.json` on Linux, `~/Library/Application Support/Claude/claude_desktop_config.json` on MacOS, or `%APPDATA%\Claude\claude_desktop_config.json` on Windows):

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

**Claude Desktop on Windows:** If the server appears in the list but never connects (no handshake, tools missing), Claude may be using its built-in Node.js runtime, which does not see packages installed with your system `npm` (including a global `mcp-remote`). Set `isUsingBuiltInNodeForMcp` to `false` at the **root** of the same config file so `npx` uses your installed Node — then restart Claude Desktop:

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

See [issue #69](https://github.com/datagouv/datagouv-mcp/issues/69) for background.

### Cursor

Cursor supports MCP servers through its settings. To configure the server:

1. Open Cursor Settings
2. Search for "MCP" or "Model Context Protocol"
3. Add a new MCP server with the following configuration:

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

### Gemini CLI

Add the following to your `~/.gemini/settings.json` file (Linux: `~/.gemini/settings.json`, MacOS: `~/.gemini/settings.json`, Windows: `%USERPROFILE%\.gemini\settings.json`):

```json
{
  "mcpServers": {
    "datagouv": {
      "httpUrl": "https://mcp.data.gouv.fr/mcp"
    }
  }
}
```

### HuggingChat

1. **Access Settings:** In the chat interface, click the + icon, select `MCP Servers`, and click `Manage MCP Servers`.
2. **Add Server:** Click the + `Add Server` button in the server management window.
3. **Configure the Server:** Enter a **Server Name** (e.g., "Data Gouv") and set the **Server URL** to `https://mcp.data.gouv.fr/mcp`. Click `Add Server` to save.
4. **Verify Connection:** Click the `Health Check` button on the new server card to confirm it displays as **Connected**. Ensure the toggle is activated to use the tools in your chat.

### IBM Bob

IBM Bob supports MCP servers through its settings. To configure the server:

1. Click the setting icon in the Bob panel.
2. Select the MCP tab.
3. Click the appropriate button:
  - Edit Global MCP: Opens the global `mcp_settings.json` file
  - Edit Project MCP: Opens the project-specific `.bob/mcp.json` file (Bob creates it if it does not exist)

Both files use JSON format with an mcpServers object containing named server configurations.

```json
{
  "mcpServers": {
    "datagouv": {
      "url": "https://mcp.data.gouv.fr/mcp",
      "type": "streamable-http"
    }
  }
}
```

### Kiro CLI

Add the following to `~/.kiro/settings/mcp.json` (Linux: `~/.kiro/settings/mcp.json`, MacOS: `~/.kiro/settings/mcp.json`, Windows: `%USERPROFILE%\.kiro\settings\mcp.json`):

```json
{
  "mcpServers": {
    "datagouv": {
      "url": "https://mcp.data.gouv.fr/mcp"
    }
  }
}
```

### Kiro IDE

Add the following to your Kiro MCP configuration file (`.kiro/settings/mcp.json` in your workspace, or for global config: Linux: `~/.kiro/settings/mcp.json`, MacOS: `~/.kiro/settings/mcp.json`, Windows: `%USERPROFILE%\.kiro\settings\mcp.json`):

```json
{
  "mcpServers": {
    "datagouv": {
      "url": "https://mcp.data.gouv.fr/mcp"
    }
  }
}
```

### Le Chat (Mistral)

*Available on all plans, including free.*

1. **Go to Connectors**: Open Mistral in your browser, then go to `Intelligence` > `Connectors`.
2. **Add a custom connector**: Click `Add connector` > `Custom MCP Connector`, give it a name (for example `DataGouv`), and set the server URL to `https://mcp.data.gouv.fr/mcp`.
3. **No authentication**: Leave authentication disabled.
4. **Create**: Click **Create**.

### Mistral Vibe CLI

Edit your Vibe config (default: Linux: `~/.vibe/config.toml`, MacOS: `~/.vibe/config.toml`, Windows: `%USERPROFILE%\.vibe\config.toml`) and add the MCP server:

```toml
[[mcp_servers]]
name = "datagouv"
transport = "streamable-http"
url = "https://mcp.data.gouv.fr/mcp"
```

See the full Vibe MCP options in the official docs: [MCP server configuration](https://github.com/mistralai/mistral-vibe?tab=readme-ov-file#mcp-server-configuration).

### OpenCode

Add to `opencode.json` (e.g. `~/.config/opencode/opencode.json` or your project root). Remote servers use the top-level `mcp` object with `type: "remote"`. See [OpenCode MCP servers](https://opencode.ai/docs/mcp-servers/).

```json
{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "datagouv": {
      "type": "remote",
      "url": "https://mcp.data.gouv.fr/mcp",
      "enabled": true
    }
  }
}
```

### VS Code

Add the following to your VS Code `mcp.json` file (Linux: `~/.config/Code/User/mcp.json`, MacOS: `~/Library/Application Support/Code/User/mcp.json`, Windows: `%APPDATA%\Code\User\mcp.json`). Run **MCP: Open User Configuration** from the Command Palette to open it.

```json
{
  "servers": {
    "datagouv": {
      "url": "https://mcp.data.gouv.fr/mcp",
      "type": "http"
    }
  }
}
```

### Windsurf

Add the following to your `~/.codeium/windsurf/mcp_config.json` (Linux: `~/.codeium/windsurf/mcp_config.json`, MacOS: `~/.codeium/windsurf/mcp_config.json`, Windows: `%USERPROFILE%\.codeium\windsurf\mcp_config.json`):

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

**Note:**
- The hosted endpoint is `https://mcp.data.gouv.fr/mcp`. If you run the server yourself, replace it with your own URL (see “Run locally” below for the default local endpoint).
- This MCP server only exposes read-only tools for now, so no API key is required.

## 🖥️ Run locally

### 1. Run the MCP server

Before starting, clone this repository and browse into it:

```shell
git clone git@github.com:datagouv/datagouv-mcp.git
cd datagouv-mcp
```

Docker is required for the recommended setup. Install it via [Docker Desktop](https://www.docker.com/products/docker-desktop/) or any compatible Docker Engine before continuing.

#### 🐳 With Docker (Recommended)

```shell
# With default settings (port 8000, prod environment)
docker compose up -d

# With custom environment variables
MCP_PORT=8007 DATAGOUV_API_ENV=demo LOG_LEVEL=DEBUG docker compose up -d

# Stop
docker compose down
```

**Environment variables:**
- `MCP_HOST`: host to bind to (defaults to `0.0.0.0`). Set to `127.0.0.1` for local development to follow MCP security best practices.
- `MCP_PORT`: port for the MCP HTTP server (defaults to `8000` when unset).
- `MCP_ENV`: environment name reported to Sentry (defaults to `local` when unset). Set explicitly to `prod`, `preprod`, or `demo` in your deployment.
- `DATAGOUV_API_ENV`: `prod` (default) or `demo`. This controls which data.gouv.fr environement it uses the data from (https://www.data.gouv.fr or https://demo.data.gouv.fr). By default the MCP server talks to the production data.gouv.fr. Set `DATAGOUV_API_ENV=demo` if you specifically need the demo environment.
- `LOG_LEVEL`: Python logging level for the application (defaults to `INFO`). Common values: `DEBUG`, `INFO`, `WARNING`, `ERROR`, `CRITICAL`.
- `SENTRY_DSN`: Sentry DSN to enable error and performance monitoring. Monitoring is disabled when unset.
- `SENTRY_SAMPLE_RATE`: sampling rate for Sentry traces and profiles (float `0.0`–`1.0`, defaults to `1.0`).

#### ⚙️ Manual Installation

You will need [uv](https://github.com/astral-sh/uv) to install dependencies and run the server.

1. **Install dependencies**
  ```shell
  uv sync
  ```

2. **Prepare the environment file**

  Copy the [example environment file](.env.example) to create your own `.env` file:
  ```shell
  cp .env.example .env
  ```

  Then optionally edit `.env` and set the variables that matter for your run:
  ```
  MCP_HOST=127.0.0.1  # (defaults to 0.0.0.0, use 127.0.0.1 for local dev)
  MCP_PORT=8007  # (defaults to 8000 when unset)
  MCP_ENV=local  # environment name sent to Sentry (defaults to local when unset)
  DATAGOUV_API_ENV=prod  # Allowed values: demo | prod (defaults to prod when unset)
  LOG_LEVEL=INFO  # Python log level (default: INFO)
  ```

  Load the variables with your preferred method, e.g.:
  ```shell
  set -a && source .env && set +a
  ```

3. **Start the HTTP MCP server**
  ```shell
  uv run main.py
  ```

### 2. Connect your chatbot to the local MCP server

Follow the steps in [Connect your chatbot to the MCP server](#-connect-your-chatbot-to-the-mcp-server) and simply swap the hosted URL for your local endpoint (default: `http://127.0.0.1:${MCP_PORT:-8000}/mcp`).

## 🚚 Transport support

The MCP server is built using the [official Python SDK for MCP servers and clients](https://github.com/modelcontextprotocol/python-sdk) and uses the **Streamable HTTP transport only**.

**STDIO and SSE are not supported**.

## 📋 Available Endpoints

**Streamable HTTP transport (standards-compliant):**
- `POST /mcp` - JSON-RPC messages (client → server)
- `GET /health` - Health check endpoint: runs a full MCP handshake and tool call. Returns `{"status":"ok",...}` with HTTP 200 if healthy, or `{"status":"mcp_unavailable"}` with HTTP 503 if the MCP stack is not responding correctly.

## 🛠️ Available Tools

The MCP server provides tools to interact with data.gouv.fr datasets and third-party APIs cataloged on the platform.

**Note:** data.gouv.fr exposes these third-party APIs (e.g., Adresse API, Sirene API) over HTTP under the `dataservices` resource paths; that is separate from data.gouv.fr's own internal APIs (Main/Tabular/Metrics) that power this MCP server.

### Datasets (static data files)

- **`search_datasets`** - Search for datasets by keywords. Returns datasets with metadata (title, description, organization, tags, resource count).

  Parameters: `query` (required), `page` (optional, default: 1), `page_size` (optional, default: 20, max: 100)

- **`search_organizations`** - List or search publishing organizations on data.gouv.fr. Returns trimmed rows (id, name, slug, acronym, badges, metrics, URLs).

  Parameters: `query` (optional; AND-style keyword search; omit or leave empty to browse), `page` (optional, default: 1), `page_size` (optional, default: 20, max: 100), `sort` (optional; e.g. `datasets`, `-datasets`), `badge` (optional; e.g. `public-service`, `certified`, `association`, `company`, `local-authority`), `name` (optional, exact name filter), `business_number_id` (optional).

- **`get_dataset_info`** - Get detailed information about a specific dataset (metadata, organization, tags, dates, license, etc.).

  Parameters: `dataset_id` (required)

- **`list_dataset_resources`** - List all resources (files) in a dataset with their metadata (format, size, type, URL).

  Parameters: `dataset_id` (required)

- **`get_resource_info`** - Get detailed information about a specific resource (format, size, MIME type, URL, dataset association, Tabular API availability).

  Parameters: `resource_id` (required)

- **`query_resource_data`** - Query data from a specific resource via the Tabular API. Fetches rows from a resource to answer questions.

  Parameters: `resource_id` (required), `page` (optional, default: 1), `page_size` (optional, default: 20, max: 200)

  Note: Recommended workflow: 1) Use `search_datasets` to find the dataset, 2) Use `list_dataset_resources` to see available resources, 3) Use `query_resource_data` with default `page_size` (20) to preview data structure. For small datasets (<500 rows), increase `page_size` or paginate. For large datasets (>1000 rows), continue paginating or use `get_resource_info` to retrieve the raw file URL and fetch it directly. Works for CSV/XLS resources within Tabular API size limits (CSV ≤ 100 MB, XLSX ≤ 12.5 MB).

### Third-party APIs

These tools use data.gouv.fr HTTP paths under `dataservices`; tool and parameter names match that API (`search_dataservices`, `dataservice_id`).

- **`search_dataservices`** - Search for third-party APIs cataloged on data.gouv.fr by keywords. Returns entries with metadata (title, description, organization, base API URL, tags).

  Parameters: `query` (required), `page` (optional, default: 1), `page_size` (optional, default: 20, max: 100)

- **`get_dataservice_info`** - Get detailed metadata for one third-party API (title, description, organization, base API URL, OpenAPI spec URL, license, dates, related datasets).

  Parameters: `dataservice_id` (required) — same as in the data.gouv.fr API and as the `id` from search results.

- **`get_dataservice_openapi_spec`** - Fetch and summarize the OpenAPI/Swagger specification for a third-party API. Returns a concise overview of available endpoints with their parameters.

  Parameters: `dataservice_id` (required)

  Note: Recommended workflow: 1) Use `search_dataservices` to find the API, 2) Use `get_dataservice_info` for metadata and documentation URL, 3) Use `get_dataservice_openapi_spec` for endpoints and parameters, 4) Call the API using the `base_api_url` per the spec.

### Metrics

- **`get_metrics`** - Get metrics (visits, downloads) for a dataset and/or a resource.

  Parameters: `dataset_id` (optional), `resource_id` (optional), `limit` (optional, default: 12, max: 100)

  Returns monthly statistics including visits and downloads, sorted by month in descending order (most recent first). At least one of `dataset_id` or `resource_id` must be provided. **Note:** This tool only works with the production environment (`DATAGOUV_API_ENV=prod`). The Metrics API does not have a demo/preprod environment.

## 🧪 Tests

### ✅ Automated Tests with pytest

Run the tests with pytest (these cover helper modules; the MCP server wiring is best exercised via the MCP Inspector):

```shell
# Run all tests
uv run pytest

# Run with verbose output
uv run pytest -v

# Run specific test file
uv run pytest tests/test_tabular_api.py

# Run with custom resource ID
RESOURCE_ID=3b6b2281-b9d9-4959-ae9d-c2c166dff118 uv run pytest tests/test_tabular_api.py

# Run with prod environment
DATAGOUV_API_ENV=prod uv run pytest
```

### 🔥 Stress Tests

Stress tests send many concurrent requests against a running MCP server. They require a running server and make real HTTP requests, so they are excluded from default `pytest` runs.

```shell
# Start the server first, then in another terminal:
uv run pytest -m stress
```

Currently includes a test that mixes normal requests with abrupt client TCP disconnects, verifying the server stays healthy and keeps serving despite the disruption. It uses `MCP_PORT` (default: `8000`) to connect to the local server.

### 🩺 Run a Health Check from the CLI

Runs a full MCP handshake and calls `search_datasets` to validate end-to-end stack health. Requires a running server and is excluded from default `pytest` runs.

```shell
# Start the server first, then in another terminal:
uv run pytest -m health_check
```

### 🛠️ Local Tool Testing Script

`scripts/call_tool.py` lets you call any MCP tool directly without manually managing the curl handshake. Requires a running server.

```shell
# Start the server first, then in another terminal:
python scripts/call_tool.py search_datasets '{"query": "IRVE"}'
python scripts/call_tool.py get_resource_info '{"resource_id": "<id>"}'
```

### 🔍 Interactive Testing with MCP Inspector

Use the official [MCP Inspector](https://modelcontextprotocol.io/docs/tools/inspector) to interactively test the server tools and resources.

Prerequisites:
- Node.js with `npx` available

Steps:
1. Start the MCP server (see above)
2. In another terminal, launch the inspector:
   ```shell
   npx @modelcontextprotocol/inspector --http-url "http://127.0.0.1:${MCP_PORT}/mcp"
   ```
   Adjust the URL if you exposed the server on another host/port.

## 🤝 Contributing

We welcome contributions! To keep the project stable and reviews manageable, please observe these rules before submitting:

- **Human review and accountability:** **Issues and pull requests** must not be raw, unreviewed AI output. You must have read, fully understood, and (for code) tested what you submit. **By opening an issue or a pull request, you certify you could explain and defend it in review without relying on an AI assistant.**
- **Keep it small:** We strictly follow a **1 feature = 1 PR** workflow.
- **Conventional commits:** Use the [Conventional Commits](https://www.conventionalcommits.org/) format for **git commit messages** and **PR titles** (e.g. `feat: add dataset search`, `fix: handle empty API response`). See the specification for allowed types, scopes, and breaking-change markers.

We use a standard review-and-deploy process:

1. **Submit a PR:** Propose your changes via a Pull Request against the `main` branch.
2. **Continuous integration:** CI runs automatically on the pull request. **All required checks must pass** before the PR can be merged (tests, linting, formatting, and type checking). Run the same checks locally—tests per [Tests](#-tests), and lint/format/type via [Code linting and formatting](#-code-linting-and-formatting) or the [pre-commit hook](#-pre-commit-hooks)—to avoid surprise CI failures.
3. **Review:** All PRs must be reviewed and approved by a maintainer before merging.
4. **Deployment process:** Once merged into `main`, maintainers deploy changes periodically to **[pre-production](https://mcp.preprod.data.gouv.fr/)** for more tests and validation before wider release.

### 🧹 Code Linting and Formatting

This project follows PEP 8 style guidelines using [Ruff](https://astral.sh/ruff/) for linting and formatting, and [ty](https://docs.astral.sh/ty/) for type checking.

**Either running these commands manually or [installing the pre-commit hook](#-pre-commit-hooks) is required before submitting contributions.**

```shell
# Lint (including import sorting) and format code
uv run ruff check --fix && uv run ruff format

# Type check (ty)
uv run ty check
```

### 🔗 Pre-commit Hooks

This repository uses a [pre-commit](https://pre-commit.com/) hook which lint and format code before each commit. Installing the pre-commit hook is strongly recommended so the checks run automatically.

**Install pre-commit hooks:**
```shell
uv run pre-commit install
```
The pre-commit hook that automatically:
- Check YAML syntax
- Fix end-of-file issues
- Remove trailing whitespace
- Check for large files
- Run Ruff linting and formatting

### 🏷️ Releases and versioning

The release process uses the [`tag_version.sh`](tag_version.sh) script to create git tags, GitHub releases and update [CHANGELOG.md](CHANGELOG.md) automatically. Package version numbers are automatically derived from git tags using [setuptools_scm](https://github.com/pypa/setuptools_scm), so no manual version updates are needed in `pyproject.toml`.

**Prerequisites**: [GitHub CLI](https://cli.github.com/) must be installed and authenticated, and you must be on the main branch with a clean working directory.

```shell
# Create a new release
./tag_version.sh <version>

# Example
./tag_version.sh 2.5.0

# Dry run to see what would happen
./tag_version.sh 2.5.0 --dry-run
```

The script automatically:
- Extracts commits since the last tag and formats them for CHANGELOG.md
- Identifies breaking changes (commits with `!:` in the subject)
- Creates a git tag and pushes it to the remote repository
- Creates a GitHub release with the changelog content

## 📄 License

This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.