Enables MCP clients to query and diagnose machine-learning training runs through read-only tools for listing experiments, reading configurations and final metrics, comparing runs, sampling metric curves, and detecting training failures.