← Projects · MCP Protocol · MLOps · PyPI · Open Source

MLOps MCP Server

FastMCP · MLflow · DVC · Git · Pandera · SciPy · Python 3.11+

An open-source MCP server that gives AI assistants direct, tool-based access to MLOps workflows — experiment tracking, model registry, dataset management, pipeline orchestration, and lineage tracing. Built, published, and maintained by me; installable by anyone.

PyPI ↗GitHub ↗
pip install mlops-mcp-server   # v0.1.0 · MIT · Python 3.11+

Problem

AI coding assistants can write training code, but they’re blind to the MLOps state around it — which experiment run is best, what version a model is at, whether a dataset has drifted, whether a pipeline DAG is even valid. Working with MLflow, DVC, and Git meant constant context-switching and copy-paste between the assistant and the tooling. This server closes that gap: the assistant calls typed MCP tools and operates on the ML workflow directly.

My role

Sole author — designed, built, documented, and published the package end-to-end.

Architecture

AI assistantClaude · any MCP host mlops-mcp-servertwo-tier tool registry:core tools always on+ 15 on-demand modules Domain modules experiments · runs, params, metrics, compare registry · register, promote, tag, deprecate datasets · profile, validate, drift (KS-test) pipelines · DAG validate, Kahn cycle detection lineage · provenance, BFS trace, Mermaid viz scaffolding · 7 ML project templates docs · model cards, dataset cards, reports env · import scan, requirements, conda Wrapped backendsMLflow · DVC · Git · filesystem Two-tier registry keeps the assistant’s context window lean: only core tools load by default; domain modules attach on demand.

Tech stack

FastMCPMCP ProtocolMLflowDVCGitPanderaSciPyPython 3.11+MIT

Key design decisions & trade-offs

  • Two-tier tool registry. An MCP server that exposes everything at once floods the model’s context with tool schemas. Core tools are always available; 15 domain modules load on demand — the assistant discovers depth only when it needs it.
  • Real algorithms, not wrappers only. Dataset drift detection uses the Kolmogorov–Smirnov test; pipeline validation runs Kahn’s algorithm for DAG cycle detection; lineage tracing is BFS over the provenance graph, rendered as Mermaid diagrams the assistant can show inline.
  • Optional-dependency extras. MLflow, Pandera, pyarrow and friends are opt-in extras — the base install stays light, and features degrade gracefully when a backend isn’t present.
  • Published, versioned, licensed. v0.1.0 on PyPI under MIT — anyone can install, inspect, and extend it.

Results

Published on PyPI (v0.1.0, MIT) · pip-installable by anyone · integrates MLflow + DVC + Git · 7 project templates