Skore MCP server

Community MCP server for Skore — not affiliated with Probabl. It exposes MCP tools for projects, evaluate / compare / train_test_split, global configuration, and a skore_report_call dispatcher that covers the report surface documented in the Skore API (e.g. ComparisonReport.get_predictions, metrics, inspection, data.analyze). Call skore_api_catalog for the exact allowlisted method names per accessor.

Repository: github.com/fanfcorp/skore-mcp

Requirements

• Python 3.10+ (same as Skore 0.15)

• Dependencies: skore, mcp[cli], pandas (see pyproject.toml)

Install

From PyPI (after the package is published):

python3.12 -m venv .venv
source .venv/bin/activate
pip install skore-mcp

From GitHub (any time, no PyPI needed):

pip install "skore-mcp @ git+https://github.com/fanfcorp/skore-mcp.git"

From a local clone (editable for development):

cd "/path/to/skore-mcp"
python3.12 -m venv .venv
source .venv/bin/activate
pip install -e .

After install, run the server as skore-mcp or python -m skore_mcp (stdio MCP).

Cursor / Claude Desktop

Add a stdio server (use your real paths):

{
  "mcpServers": {
    "skore": {
      "command": "/path/to/MCP Skore/.venv/bin/python",
      "args": ["-m", "skore_mcp"],
      "env": {
        "SKORE_WORKSPACE": "/path/to/your/skore/workspace"
      }
    }
  }
}

Alternatively, after pip install into an environment:

{
  "mcpServers": {
    "skore": {
      "command": "skore-mcp"
    }
  }
}

Tools

Session keys

skore_evaluate_csv always registers the report in-memory and returns session_report_key. Use that with skore_report_call (and skore_compare_session_reports) without persisting to a project. Persisted reports use project_name + report_id (id from summarize) instead.

API coverage

Skore’s Python API is large (many classes and plot objects). This server maps one tool (skore_report_call) to the methods Skore exposes on reports and accessors, with an explicit allowlist so behavior stays predictable. Plots are returned as base64 PNG and tables as CSV in JSON when serialization supports it. It is not a line-for-line duplicate of every overload in the docs, but it covers the public patterns for reports, metrics, inspection, and data analysis.

Hub projects use project_name like workspace_slug/project_slug per Skore docs. Call skore_hub_login (or your hub plugin’s env vars) before hub operations.

Notes

• On SQLite 3.41+, the server applies a small compatibility patch for diskcache (used by local Skore storage) so local projects work on current macOS/Python builds.

• Logs go to stderr; do not print to stdout when using stdio MCP.

License

This project is licensed under the MIT License — see LICENSE. Skore and other dependencies remain under their own licenses.

Publish or clone from GitHub

PyPI via Trusted Publishing (recommended)

The repo includes .github/workflows/publish.yml. On PyPI → skore-mcp → Publishing (or when creating the project), add a pending publisher:

Then merge the workflow on main, bump the version in pyproject.toml / skore_mcp/__init__.py, and create a GitHub Release (or run the workflow manually with Actions → Publish to PyPI → Run workflow). The release event triggers the upload; no long-lived PYPI_API_TOKEN is required.

Clone

After cloning:

git clone https://github.com/fanfcorp/skore-mcp.git
cd skore-mcp
python3.12 -m venv .venv && source .venv/bin/activate
pip install -e .

To push updates (replace the remote URL if your fork or username differs):

git remote add origin https://github.com/fanfcorp/skore-mcp.git
git branch -M main
git push -u origin main

Or create the repo and push in one step (with GitHub CLI authenticated):

gh repo create fanfcorp/skore-mcp --public --source=. --remote=origin --push