Summary
Experiment history and the registry, answerable in a sentence.
The read half covers the question people actually have — which run, what did it score, what is registered — without opening the MLflow UI. The write half reaches into the registry: creating versions from a run's artifacts, moving aliases, and deleting models, which is promotion, not just bookkeeping.
What it is
An MCP server exposing MLflow experiment tracking and model registry operations to an MCP client, published on PyPI as mlflow-mcp-server.
What you get
- Experiments fetched by ID or by name, and searched with name matching and pagination
- A run's full detail — metrics, parameters, tags, and whether it is a parent, child or standalone run
- Runs listed for an experiment with pagination, so a long-running sweep is readable
- Registered models searched and listed, with model versions browsed and filtered
- New registered models created with description and tags, and new versions created from a run's artifacts
- Registered models renamed, and versions deleted
- Aliases assigned to a model version, which is how a stage pointer gets moved
- Registered models deleted when they should no longer exist
Requirements
A reachable MLflow tracking server, set through MLFLOW_TRACKING_URI. It runs with `uvx mlflow-mcp-server`, so no clone is needed. Whatever authentication your tracking server enforces still applies.
Setup effort
One command — uvx mlflow-mcp-server