MLAgent

MLAgent profiles a tabular dataset, tests a small set of scikit-learn pipelines, and packages the best cross-validated predictor. An LLM may propose experiments, but deterministic code validates each plan, runs evaluation, and selects the winner.

Install

Python 3.10 or newer is required.

uv sync

Run

Start with the offline planner to check the full pipeline without an account:

uv run mlagent run demo:breast_cancer --target target --planner fake

Use a logged-in Codex or Claude Code subscription:

codex login
uv run mlagent run data.csv --target outcome --planner codex

claude auth login --claudeai
uv run mlagent run data.csv --target outcome --planner claude

Or use the OpenAI API:

export OPENAI_API_KEY="..."
uv run mlagent run data.csv --target outcome --planner llm

The model receives column names, summary statistics, and previous experiment results. It does not receive dataset rows. Both CLI planners run in an empty temporary directory; Claude's tools are disabled and Codex uses its read-only sandbox.

Each run writes a predictor, schema, leaderboard, run record, and Markdown report to runs/<timestamp>/ by default.

Predict

uv run mlagent predict runs/<timestamp> new_rows.csv --out predictions.csv

Test

uv run pytest -q