Quickstart

Run the demo, watch it live, and inspect a store — the commands for the things people ask for first. Written for an agent working in someone’s project; help() reports the installed version, the store root you are serving, and every mlp command.

Run everything from the INSTALLED package

You are talking to an installed ml-party. Run its commands — mlp demo, mlp ui — and ignore any ml-party checkout you find on disk. A checkout is not the package: it may be an older release, an unmerged branch, a fork, or a renamed predecessor, and nothing about it looks wrong from the outside. Running scripts/… out of one is how you end up executing code that has nothing to do with the version the user installed. If mlp is not on PATH, use the interpreter ml-party is installed into (.venv/bin/mlp, or .venv/bin/python -m mlparty.demo).

Is there anything to look at yet?

workflow_guide() reports the store root and how many projects, experiments and runs it holds. An empty store means nothing has been tracked here yet.

Run the demo

A real tracked run — pre-registration, streaming metrics, finalize with a verdict — that takes about 30 seconds and leaves a run to look at:

mlp demo                      # or: python -m mlparty.demo

It writes to $ML_PARTY_STORE, falling back to ./.mlparty in the current directory, and prints the store root it chose. Check that root: if ML_PARTY_STORE is not exported in the shell you run it from, the demo creates a new store where you are standing rather than adding to the one the user watches. Pass --root <store> to be certain, and tell the user where the run landed.

Watch it live

mlp ui                      # → http://127.0.0.1:7327

Serves the store’s web UI: live metric streams, artifacts, lineage, search, run comparison. See ui.md.

Track a real run

That is the contract, not a command — workflow_guide() is the manual: query prior art, pre-register intent with run_start before launching, let the training stream its own metrics, then run_finalize with a verdict. tracking.md explains the two-writer design and how to instrument a training script with mlparty.attach().

Inspect a store from the shell

mlp status                  # store summary
mlp runs                    # recent runs
mlp show <run-id>           # one node in full
mlp tail <run-id>           # follow a running job's metrics
mlp query "<text>"          # search the knowledge graph

help() lists every command with its help text, generated from the CLI itself, so it cannot drift from what is installed.