MLflow Experiment Tracking AI Skill Guide (Claude)

Overview & Engine Architecture

MLflow Tracking records experiments → runs with parameters, metrics, tags, and artifacts. A tracking server (or local mlruns/) holds metadata; an artifact store (local, S3/MinIO, or server-proxied) holds files. The Model Registry versions models and (since stages are deprecated) uses aliases (champion, challenger) and tags for deployment intent. Agents log enough to reproduce a run, align client/server tracking URIs, and promote only evaluated candidates via aliases—not every train job.