MLflow MCP Server

A Model Context Protocol (MCP) server that enables LLMs to interact with MLflow tracking servers. Query experiments, analyze runs, compare metrics, manage the model registry, and promote models to production — all through natural language.

Features

• Experiment Management: List, search, and filter experiments
• Run Analysis: Query runs, compare metrics, find best performing models
• Metrics & Parameters: Get metric histories, compare parameters across runs
• Artifacts: Browse and download run artifacts
• LoggedModel Support: Search and retrieve MLflow 3 LoggedModel entities
• Model Registry: Full registry management — register, tag, alias, stage, and promote models
• Write & Delete Actions: Tag, alias, register, promote, and delete runs/experiments/models
• MCP Prompts: Built-in guided workflows for common tasks
• Pagination: Offset-based pagination for browsing large result sets

Tools

Experiments

Runs

Metrics & Parameters

Artifacts

Analysis & Comparison

Logged Models (MLflow 3)

Model Registry

Health

Prompts

Built-in guided workflows available as slash commands in Claude:

Debugging

Use MCP Inspector to browse tools, call them with custom inputs, and inspect raw responses — without involving an LLM.

Published package:

npx @modelcontextprotocol/inspector uvx mlflow-mcp

Local source:

npx @modelcontextprotocol/inspector uv run --project /path/to/mlflow-mcp mlflow-mcp

Set MLFLOW_TRACKING_URI in the Inspector's environment panel, or pass it inline:

MLFLOW_TRACKING_URI=http://127.0.0.1:5000 npx @modelcontextprotocol/inspector uvx mlflow-mcp

Links

• PyPI Package
• GitHub Repository
• MLflow Documentation
• Model Context Protocol

Installation

Using uvx (Recommended)

# Run directly without installation
uvx mlflow-mcp

# Or install globally
pip install mlflow-mcp

From Source

git clone https://github.com/kkruglik/mlflow-mcp.git
cd mlflow-mcp
uv sync
uv run mlflow-mcp

Configuration

Claude Desktop

Add to your Claude Desktop config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/claude/claude_desktop_config.json

{
  "mcpServers": {
    "mlflow": {
      "command": "uvx",
      "args": ["mlflow-mcp"],
      "env": {
        "MLFLOW_TRACKING_URI": "http://localhost:5000"
      }
    }
  }
}

Claude Code (project-scoped)

Add .mcp.json to your project root:

{
  "mcpServers": {
    "mlflow": {
      "command": "uvx",
      "args": ["mlflow-mcp"],
      "env": {
        "MLFLOW_TRACKING_URI": "http://localhost:5000"
      }
    }
  }
}

Authenticated Server

For MLflow servers with authentication, add credentials to the env block:

{
  "mcpServers": {
    "mlflow": {
      "command": "uvx",
      "args": ["mlflow-mcp"],
      "env": {
        "MLFLOW_TRACKING_URI": "https://mlflow.company.com",
        "MLFLOW_TRACKING_USERNAME": "your-username",
        "MLFLOW_TRACKING_PASSWORD": "your-password"
      }
    }
  }
}

For Databricks or token-based auth, use MLFLOW_TRACKING_TOKEN instead:

{
  "mcpServers": {
    "mlflow": {
      "command": "uvx",
      "args": ["mlflow-mcp"],
      "env": {
        "MLFLOW_TRACKING_URI": "https://mlflow.company.com",
        "MLFLOW_TRACKING_TOKEN": "your-token"
      }
    }
  }
}

Environment Variables

Usage Examples

Explore experiments and runs

"Show me all experiments. Which ones were updated recently?"

"What metrics and parameters are tracked in experiment 'fraud-detection'?"

"Get the top 10 runs in 'fraud-detection' sorted by test/f1. Show me the params that differ most between the top 3."

"Find all runs tagged with model_type=lightgbm and compare their recall scores."

Analyze a training run

"Show me the full details of run abc123 — metrics, params, and artifacts."

"Plot the training loss curve for run abc123." (Claude fetches metric history and renders a chart)

"This run has a parent — show me the parent run and compare their metrics."

Find and register the best model

"Find the best logged model in experiment 'fraud-detection' by test/recall. Register it as 'fraud-classifier' with a selection_metric tag."

"Which logged model in experiments 1 and 2 has the highest F1 score on the validation set?"

"Register the model from run abc123 artifact path 'model/' as 'my-classifier'."

Manage the model registry

"Show me all versions of 'fraud-classifier' with their aliases and stages."

"Set the champion alias on version 3 of fraud-classifier."

"Update the description of fraud-classifier v3 to explain what dataset it was trained on."

"Copy fraud-classifier v3 to a separate 'fraud-classifier-prod' model as the production entry."

Audit your MLflow setup

"Audit my MLflow setup"

(Triggers the audit_mlflow_setup built-in prompt — Claude explores experiments, runs, artifacts, and the model registry, then scores each area against Google/Databricks best practices)

| Category             | Score  | Top Issue                                      |
|----------------------|--------|------------------------------------------------|
| Experiment Org       |  5/10  | Flat namespace, no dot-notation hierarchy      |
| Parameter Logging    |  7/10  | No parent-child nesting for tuning sweeps      |
| Metric Logging       |  6/10  | Only final values logged, no training curves   |
| Tagging Strategy     |  5/10  | Params duplicated as tags; stale test_tag      |
| Artifact Management  |  2/10  | No log_model(); artifacts on local disk        |
| Model Registry       |  3/10  | Duplicate prod models instead of aliases       |
| Reproducibility      |  3/10  | No git SHA; no mlflow.log_input() datasets     |
| Mean Score           |  4.4/10|                                                |

Top 3 improvements:
1. Call log_model() and move artifact store to S3/GCS
2. Add git SHA tag + mlflow.log_input() for dataset tracking
3. Consolidate registry to one model entry with @champion alias

End-to-end promotion workflow

"Find the best model in 'fraud-detection' by test/recall, register it as 'fraud-classifier', tag it with the framework and problem type, and set it as champion. Ask me before copying to prod."

(This maps directly to the promote_best_model built-in prompt)

Requirements

• Python >=3.10
• MLflow >=3.4.0
• Access to an MLflow tracking server

License

MIT License - see LICENSE file for details.