MLflow

You run an MCP server that exposes MLflow capabilities through a standard interface, letting you browse experiments and runs, inspect metrics and parameters, and query registered models from your AI assistant or automation tooling. This guide walks you through practical setup, usage patterns, and configuration so you can start querying MLflow data via MCP right away.

How to use

Start the server and expose MLflow data to your AI assistant. Once the MCP server is running, your assistant can perform common MLflow tasks such as locating experiments by name, retrieving detailed run information with metrics and parameters, listing all runs for a given experiment, and browsing registered models and their versions. Use natural language prompts like: “Show me experiments containing 'customer' in the name,” “Get details for run abc123 with metrics,” or “List the latest versions of registered models.” You can chain these capabilities to build workflows that monitor experiments, compare runs, or surface model lineage to users.

How to install

Prerequisites and install steps are shown here so you can get up and running quickly.

# Prerequisites
python3 --version
uv --version

# Install dependencies and start the MCP server
# 1) Install uv (package manager used here)
# 2) Run the server (see the configuration below for details)

Configuration

Configure how the MCP server connects to MLflow and how you expose the server to your MCP client.

{
  "mcpServers": {
    "mlflow": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/yesid-lopez/mlflow-mcp-server", "mlflow_mcp_server"],
      "env": {
        "MLFLOW_TRACKING_URI": "YOUR_TRACKING_URI"
      }
    }
  }
}

Security and access

Limit access to the MCP server to trusted clients and keep the MLflow tracking URI secure. If you expose the server over a network, implement authentication and network-level protections to prevent unauthorized access to MLflow data.

Notes and tips

• The server uses the MLflow Python client to fetch experiments, runs, and models. Ensure your MLflow instance is reachable from the MCP server host via the tracking URI you provide.

• If you need to point to a different MLflow instance, adjust the MLFLOW_TRACKING_URI environment variable in the configuration snippet above.

Troubleshooting

• If the server fails to start, verify that uvx is installed and the tracking URI is correct and reachable.

• Check connectivity from the MCP host to the MLflow tracking server and ensure any required TLS or firewall rules allow traffic.

Example usage scenarios

• Find all experiments with a name containing a keyword, e.g., “reporting”.

• Retrieve a run by ID and inspect its metrics, parameters, and tags.

• List registered models and inspect their versions to see latest updates.

Available tools

get_experiment

Retrieve experiment details by ID.

get_experiment_by_name

Find experiments by name.

search_experiments

List and filter experiments with optional pagination.

get_run

Retrieve detailed run information including metrics, parameters, and tags.

get_experiment_runs

List all runs for a specific experiment.

get_registered_models

Search and list registered models.

get_model_versions

Browse model versions with filtering capabilities.