MCP for AI agents

Install the MCP extra, then register the breezeml-mcp server with your agent. The console script speaks stdio transport, so one line wires it into Claude Code. Once connected, the agent sees eleven tools: inspect_data, audit, train, compare, predict, explain, report, model_card, export, deploy, and save.

pip install "breezeml[mcp]"

# register the server with Claude Code
claude mcp add breezeml -- breezeml-mcp

Look before you leap. inspect_data profiles a CSV (rows, features, missing values, class balance) and audit scans it for target leakage, ID columns, constants, duplicates, and label noise. Both return JSON so the agent can reason about whether the data is even worth training on.

// tool: inspect_data
{ "csv_path": "churn.csv", "target": "churn" }

// result (abridged)
{ "rows": 7043, "features": 19, "missing": 0,
  "target": "churn", "classes": { "No": 5174, "Yes": 1869 } }

Trains the auto-selected model and stores it in the session under a model_id the agent reuses for later calls. The result includes the task, the chosen estimator, the cross-validated report, and a plain-English list of every pipeline decision.

// tool: train
{ "csv_path": "churn.csv", "target": "churn" }

// result
{ "model_id": "model_1", "task": "classification",
  "estimator": "HistGradientBoostingClassifier",
  "report": { "f1": 0.81, "roc_auc": 0.86 },
  "decisions": ["detected classification from 2 classes", "..."] }

Benchmarks every built-in model on the same cross-validation folds and returns a ranked leaderboard, so the agent picks a winner fairly instead of guessing an algorithm.

// tool: compare
{ "csv_path": "churn.csv", "target": "churn" }

// result
{ "task": "classification",
  "leaderboard": [
    { "model": "hist_gradient_boosting", "f1": 0.81 },
    { "model": "random_forest", "f1": 0.79 },
    { "model": "logistic", "f1": 0.76 } ] }

The gate. Runs the full honesty gauntlet on a trained model and returns a single SHIP, WARN, or STOP verdict with reasons. Agents are instructed to call this and confirm SHIP before they deploy or export. Pass a sensitive column to add a fairness check.

// tool: report
{ "model_id": "model_1", "csv_path": "churn.csv", "target": "churn" }

// result
{ "verdict": "SHIP", "target": "churn", "task": "classification",
  "reasons": [],
  "sections": { "performance": { "beats_baseline": true } } }

The shipping tools. deploy writes a complete FastAPI plus Docker serving directory, export writes a standalone scikit-learn script, model_card writes a Markdown governance doc, and save persists the model to a .joblib file. An honest agent only reaches for these after report returns SHIP.

// tool: deploy (only after a SHIP verdict)
{ "model_id": "model_1", "out_dir": "api/" }

// result
"Serving app written to api/ (run: cd api/ && pip install -r requirements.txt && uvicorn app:app)"

The whole point: the server nudges agents toward a sound workflow. Inspect the data, compare models, train the winner, then run report. Only if the verdict is SHIP does the agent deploy. A WARN or STOP stops the line and the agent reports back instead of shipping a broken model.

Agent: inspect_data(churn.csv, churn)   -> looks clean, 2 classes
Agent: compare(churn.csv, churn)        -> hist_gradient_boosting wins
Agent: train(churn.csv, churn)          -> model_1, f1 0.81
Agent: report(model_1, churn.csv, churn)-> verdict SHIP
Agent: SHIP confirmed, proceeding.
Agent: deploy(model_1, api/)            -> FastAPI app written