Creativity · MCP — server

MCP MLflow Server

The MCP MLflow Server wraps the MLflow REST API for tracking and model registry. MCP clients can search experiments, read run artifacts and metrics, register a model version, and transition between Staging and Production. It's a natural companion to a CI pipeline that promotes the best model of the week.

MCP facts

Kind: server
Ecosystem: anthropic-mcp
Language: Python (mlflow)
Transports: stdio

Capabilities

Tools: search_runs, get_run, register_model, transition_model_version, list_registered_models
Resources: mlflow://run/{id}, mlflow://model/{name}/version/{version}
Auth: MLflow tracking URI + optional Databricks PAT or basic auth

Install

pipx install mcp-server-mlflow

Configuration

{
  "mcpServers": {
    "mlflow": {
      "command": "mcp-server-mlflow",
      "env": {
        "MLFLOW_TRACKING_URI": "https://mlflow.example.com",
        "MLFLOW_TRACKING_TOKEN": "${MLFLOW_TOKEN}"
      }
    }
  }
}

Frequently asked questions

Can it work with Databricks-managed MLflow?

Yes — set MLFLOW_TRACKING_URI to databricks and pass a Databricks personal access token.

Is stage transition safe from an LLM?

Always gate promotions behind human approval, CI checks, or an evaluation tool. MLflow transitions are permissioned, but an LLM alone should not promote to Production.

Does it expose served endpoints?

Optionally — an invoke_model tool can send a payload to a deployed MLflow serving endpoint. Treat this like any other inference tool.

Sources

MLflow REST API — accessed 2026-04-20
Model Context Protocol — accessed 2026-04-20
MCP servers repo — accessed 2026-04-20