mcp-ml-lab

Let AI agents run real ML experiments end-to-end.

An MCP server that gives Claude (or any MCP-aware AI agent) the ability to profile a CSV, define an ML task, tune XGBoost and LightGBM with Optuna, and produce a markdown report with feature importance — all from natural language.

Why this exists

The existing ML-related MCP servers wrap MLflow, ZenML, or Weights & Biases and expose them as read-only — agents can browse experiment history but can't actually run anything. mcp-ml-lab fills the gap: it lets agents execute the full experimentation loop from a user's natural-language request.

A user typing "train a model on titanic.csv to predict survival" should not need to know what XGBoost is, what cross-validation is, or how to write a hyperparameter search. The agent handles all of that — mcp-ml-lab is the tools layer that makes it possible.

Quick start

pip install mcp-ml-lab

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "ml-lab": {
      "command": "mcp-ml-lab"
    }
  }
}

Restart Claude Desktop. The five tools below are now available.

Example queries

Try these in Claude Desktop with mcp-ml-lab connected:

• "Profile this CSV and tell me if there's class imbalance"
• "Compare XGBoost and LightGBM on titanic.csv with 60 seconds of tuning"
• "Show me the top 10 features the winning model used"
• "How did my last three experiments on the wine dataset compare?"

Tools

Tool	What it does
inspect_data	Profile a CSV — shape, dtypes, nulls, summary stats, class balance
define_task	Register an ML task (CSV + target + classification/regression)
run_experiment	Train one or more models, optionally tuning with Optuna
get_results	Markdown report with metrics, hyperparameters, feature importance
compare_runs	Side-by-side comparison of multiple experiments

Each tool's full signature is in its docstring; they self-document to the LLM.

How it works

Claude Desktop  ───MCP/stdio───  mcp-ml-lab server
                                  │
                                  ├── data.py        CSV loading, schema inference, preprocessor
                                  ├── trainers/      Pluggable XGBoost + LightGBM adapters
                                  ├── search.py      Stratified CV + Optuna TPE tuning
                                  ├── metrics.py     Accuracy, F1, AUC, log loss
                                  ├── storage.py     SQLite via SQLAlchemy 2.0
                                  └── reporting.py   Markdown report generation

All experiments and trials are persisted to ~/.mcp-ml-lab/store.db so an agent can refer back to runs across sessions.

Full design notes in ARCHITECTURE.md.

Roadmap

v0.1.0 ships classification with XGBoost and LightGBM. Planned for v0.2.0+:

• Regression tasks
• Time series forecasting (sktime / darts integration)
• Deep learning baselines (pytorch-tabular)
• Optuna multi-objective search (accuracy × latency × model size)
• Persisted model artifacts with Docker reproducibility
• Permutation feature importance (bias-free alternative to gain importance)
• Notebook export — emit a Jupyter notebook that reproduces the winning run

Issues and PRs welcome.

Development

git clone https://github.com/rohithraju-ops/mcp-ml-lab.git
cd mcp-ml-lab
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest -v

Local debugging is easiest with the MCP Inspector:

npx @modelcontextprotocol/inspector mcp-ml-lab

License

MIT.