Mcp Ml Lab

If you want Claude to actually run ML experiments instead of just browsing past results, this is the missing piece. It exposes five tools that let agents profile CSVs, define classification tasks, tune XGBoost and LightGBM models with Optuna, and generate markdown reports with feature importance. Everything persists to SQLite so the agent can compare runs across sessions. You tell Claude "train a model on titanic.csv to predict survival" and it handles the full loop: preprocessing, stratified cross validation, hyperparameter search, evaluation. Regression and time series forecasting are on the roadmap. Useful when you want to prototype models conversationally without writing boilerplate.