Skip to main content

ADR-001: LoggedModel Operations Require Gateway MCP

Status​

Accepted (July 2026)

Context​

The official MLflow MCP server (mlflow mcp run in MLflow 3.14) does not expose LoggedModel CRUD operations. The full MLflow Python SDK supports:

  • mlflow.register_logged_model()
  • mlflow.set_logged_model_tag()
  • mlflow.search_logged_models()
  • mlflow.get_logged_model()

But none of these are available as MCP tools.

Decision​

Agent Lens skills that need LoggedModel data (agent-registry, compliance-export, audit-trail) will:

  1. Today: Derive agent identity from experiments and evaluation run tags
  2. M2: Use the Agent Lens Gateway MCP which bridges the SDK gap

The Gateway will expose:

  • search_agents (wraps search_logged_models)
  • get_agent_status (wraps get_logged_model + tags)
  • qualify_agent (wraps set_logged_model_tag + audit)

Consequences​

  • No full agent registry until Gateway ships (M2)
  • Skills document this limitation explicitly
  • Integration tests validate only official MCP tools
  • Gateway becomes the single point for SDK-bridging, not scattered workarounds