ml-party

An agent-native platform for ML experiment tracking and lineage/knowledge.

A durable, agent-legible knowledge substrate for ML work — a lab notebook that agents write (over MCP) and query, and that humans read live (CLI + web UI) — that also happens to be a full local tracker: metrics, artifacts, checkpoints, reproducibility, all stored locally.

Linux/macOS (Windows via WSL), Python ≥ 3.11. The web UI ships prebuilt in the wheel:

pip install mlparty
mlp init --root ~/.mlparty --no-mcp         # one store holds all your projects
export ML_PARTY_STORE=~/.mlparty            # put this in your shell profile

mlp demo &                                  # a real run: contract + live metrics
mlp ui                                      # → http://127.0.0.1:7327

Then register the store with your agent — one command for Claude Code, a snippet to paste for any other MCP client — and restart the session: MCP setup & tools. From then on you can ask the agent about ml-party itself: these guides ship inside the wheel and a help tool serves them, so “how do I run the demo?” or “how do I instrument my training script?” is answered from the installed package rather than from a search engine. That page also covers one store versus one per project, which is worth deciding before you accumulate runs.

To develop against a checkout instead, see From source (adds Node ≥ 20 for the UI build).

From there: the MCP server teaches agents the workflow (MCP setup & tools); training scripts attach as the second writer (Tracking runs); humans watch live in the web UI (The web UI); registered actions let agents drive jobs, audited (Run control (actions)).