mlops-mcp-server

MLOps MCP Server

<!-- -->

MCP server for common MLOps workflows

An MCP (Model Context Protocol) server that gives AI assistants like Claude direct access to MLOps workflows — experiment tracking, model registry, dataset management, pipeline orchestration, lineage tracing, and more. Wraps DVC, MLflow, and Git rather than replacing them.

---

Features

• Experiment tracking — create runs, log params/metrics/artifacts, compare runs, find best run, export to CSV
• Model registry — register, promote, tag, deprecate, and compare model versions; generate model cards
• Dataset management — profile datasets, validate schemas, detect statistical drift (KS-test), split, merge, and generate dataset cards
• Pipeline management — create, validate, and visualize DAG pipelines as Mermaid diagrams; cycle detection via Kahn's algorithm
• Data lineage — record artifact lineage, trace provenance with BFS, visualize as Mermaid graphs, check integrity
• File operations — full file/directory CRUD, content search, disk usage, batch operations
• Project scaffolding — 7 ML project templates with component injection and structure validation
• Documentation generation — model cards, dataset cards, experiment reports, pipeline docs, project READMEs, API docs
• MLflow integration — optional; wraps tracking, registry, and artifact operations
• DVC integration — optional; wraps data versioning and pipeline reproduction
• Git operations — optional; status, add, commit, log, .gitignore generation
• Environment tools — scan imports, generate requirements, check conflicts, create conda env files

Tools are registered in two tiers: a small set of always-on tools for quick file ops and discovery, and 15 domain modules you activate on demand to avoid flooding the agent context window.

---

Installation

Install directly from GitHub (available now)

pip install git+https://github.com/anant-patankar/mlops-mcp-server.git

Using uv:

uv add git+https://github.com/anant-patankar/mlops-mcp-server.git

Install from PyPI (coming soon)

# Core install
pip install mlops-mcp-server

# With optional extras
pip install mlops-mcp-server[mlflow]
pip install mlops-mcp-server[mlflow,compare,notebooks,parquet]

Using uv:

uv add mlops-mcp-server
uv add "mlops-mcp-server[mlflow,compare,notebooks,parquet]"

---

Usage

Claude Desktop

Add to your Claude Desktop MCP config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

First install the package, then add to your Claude Desktop config:

{
  "mcpServers": {
    "mlops": {
      "command": "mlops-mcp-server"
    }
  }
}

Restart Claude Desktop. The server starts automatically when Claude connects.

Example prompts

Once connected, you can ask Claude things like:

"Show me all experiment runs, find the one with the best validation accuracy, and register that model in the registry."

"Profile data/train.csv, check it against schema.yaml, then split it 70/15/15 and save to data/splits/."

"Create a pipeline with three stages: preprocess → train → evaluate, validate it for cycles, and show me the Mermaid diagram."

Running directly

mlops-mcp-server

---

Optional Dependencies

Extra	Installs	Enables
mlflow	mlflow	MLflow tracking and registry tools
compare	deepdiff	Structured diff in compare_runs
notebooks	nbformat	Notebook summary in get_notebook_summary
parquet	pyarrow	Parquet read/write in dataset tools
models	joblib, onnx, safetensors	Model file format support
validation	pandera, scipy	Schema validation and drift statistics
templates	jinja2	Jinja2 templates in create_model_card

All extras are optional — the server runs without them and returns clear errors if a missing dependency is needed.

---

License

MIT — see LICENSE.