LLM OBSERVABILITY / MLflow
MLflow MCP Server
Ships inside the mlflow package and exposes trace search, feedback and assessment logging, and optionally experiment, run, model and deployment commands, to an MCP client.
“Connect AI assistants and coding tools to MLflow using the Model Context Protocol (MCP). Search traces, analyze performance, log feedback, and more from Claude Code, Codex, Cursor, and Gemini CLI.”
01 / THE REASONING
Why this made the selection.
- The server is part of MLflow itself at this commit (mlflow/mcp/server.py and cli.py), installed with the mcp extra, and the documentation page at the same commit gives configuration for VS Code, Cursor and Claude Code.
- The documented tools include log_feedback, log_expectation and assessment updates on traces, and server.py maps MLFLOW_MCP_TOOLS to tool categories: genai (default: traces, scorers, experiments, runs), ml (experiments, runs, models, deployments) or all.
02 / THE REVIEW RECORD
What we actually inspected.
Source review has boundaries.
A clear record is more useful than a “safe” badge.
Material inspected
- docs/docs/genai/mcp/index.mdx at b033c768636ed945de47356a7f14e922ad2d9674
- mlflow/mcp/server.py and mlflow/mcp/cli.py at the same commit
- LICENSE.txt and pyproject.toml at the same commit
Our findings
- run_server builds a FastMCP server named Mlflow MCP and calls mcp.run(show_banner=False); the documented client configurations start it as a command (uv run --with mlflow[mcp]>=3.5.1 mlflow mcp run).
- pyproject.toml defines the mcp extra as fastmcp<4,>=2.7.0 and click!=8.3.0. LICENSE.txt is Apache 2.0, Copyright 2018 Databricks, Inc.
- For Databricks, the documentation sets the tracking URI to databricks with DATABRICKS_HOST and DATABRICKS_TOKEN.
Not established by this review
- pip or uv install
- Any call to an MLflow tracking server
- The ml and all tool categories
The review applies to the material and revision named here. A newer upstream release can change its behavior.
03 / PUT IT TO WORK
Connect a server deliberately.
DOCUMENTED COMMAND
pip install 'mlflow[mcp]>=3.5.1' Copying does not execute this command. It may retrieve a newer version than the reviewed source.
- Install MLflow with the mcp extra as above; the documentation's client blocks instead run it through uv. It was not run.
- Set MLFLOW_TRACKING_URI to your tracking server, and MLFLOW_MCP_TOOLS to limit tool categories.
Before you start
- Python with pip or uv
- MLflow 3.5.1 or newer
- An MLflow tracking server or Databricks workspace
Documented configuration example
Use the client and authentication context described in the setup guidance. Configuration formats vary between hosts.
{
"mcpServers": {
"mlflow-mcp": {
"command": "uv",
"args": [
"run",
"--with",
"mlflow[mcp]>=3.5.1",
"mlflow",
"mcp",
"run"
],
"env": {
"MLFLOW_TRACKING_URI": "<MLFLOW_TRACKING_URI>"
}
}
}
} 04 / FOLLOW THE EVIDENCE
The source trail.
Our notes are separate from the original resource.
Check upstream before adopting a new version.
https://github.com/mlflow/mlflow/blob/b033c768636ed945de47356a7f14e922ad2d9674/docs/docs/genai/mcp/index.mdx
Supports: Experimental status, Installation, Client configuration, Tools, Tool categories
- MLflow MCP server.py ↗Checked
https://github.com/mlflow/mlflow/blob/b033c768636ed945de47356a7f14e922ad2d9674/mlflow/mcp/server.py
Supports: Tool category mapping, FastMCP server run
- MLflow MCP cli.py ↗Checked
https://github.com/mlflow/mlflow/blob/b033c768636ed945de47356a7f14e922ad2d9674/mlflow/mcp/cli.py
Supports: mlflow mcp run command, Usage-event decorator
- MLflow LICENSE.txt ↗Checked
https://github.com/mlflow/mlflow/blob/b033c768636ed945de47356a7f14e922ad2d9674/LICENSE.txt
Supports: Apache License 2.0, Copyright 2018 Databricks, Inc.
- MLflow pyproject.toml ↗Checked
https://github.com/mlflow/mlflow/blob/b033c768636ed945de47356a7f14e922ad2d9674/pyproject.toml
Supports: mcp extra dependencies