Yes. Core MLflow is Apache 2.0 and free forever. You only pay for your own infrastructure and any LLM API calls used for evaluation or judging.

Can I use MLflow with my existing AI infrastructure?

Yes. It's framework- and cloud-agnostic, with 100+ integrations. Point it at your existing Postgres, S3, or OTel backend.

What is the MLflow Model Registry?

A centralized catalog for versioning models, tracking lineage, and managing stage transitions or aliases for deployment.

How does MLflow compare to LangSmith or Braintrust?

MLflow is open source and self-hostable with no per-seat pricing and OTel-native tracing. Proprietary tools offer more polished managed UX but lock you in.

Does MLflow support OpenTelemetry?

Yes. Tracing is built on OTel, and the server can ingest OTLP spans, so traces can be exported to any OTel-compatible backend.

How does MLflow experiment tracking work?

Each run logs params, metrics, and artifacts to a backend store (SQL) and artifact store (S3, GCS, etc.), addressed by a tracking URI.

Directory/Analytics/LLMOps/MLflow

MLflow

MLflow is the Apache 2.0 open-source platform for the full AI lifecycle: LLM/agent observability built on OpenTelemetry, systematic evaluation with 70+ built-in LLM judges and scorers, a prompt registry with lineage tracking and automated prompt optimization, an OpenAI-compatible AI Gateway for routing and cost control, and a one-command Agent Server for serving. On the classical ML side it handles experiment tracking, hyperparameter tuning, model packaging, a model registry, and deployment. It integrates with 100+ frameworks (OpenAI, Anthropic, LangChain, LangGraph, DSPy, CrewAI, LlamaIndex, PyTorch, HuggingFace) and supports Python, TypeScript, Java, R, and OpenTelemetry natively. Built for teams that need production-grade debugging and monitoring without handing their data or stack to a proprietary vendor. Deploy self-hosted with `uvx mlflow server`, or run the managed version on Databricks.

Analytics

Visit website

Added on

Nov 4, 2026

Open sourceApache 2.0LLM tracingOpenTelemetryEvaluationPrompt registryAI GatewayAgent serverExperiment trackingModel registry

Product information

Everything worth knowing about MLflow

The complete picture — from everyday features to the technical detail builders and IT folks dig for.

Core features

What it actually does

Tracing & Observability

Capture complete traces of LLM applications and agents with deep insights into behavior. Monitor production quality, costs, and safety across any LLM provider and agent framework.

Evaluation & LLM Judges

Run systematic evaluations, track quality metrics over time, and catch regressions before production. Choose from 50+ built-in metrics and LLM judges, or define custom evaluators with flexible APIs.

Prompt Registry & Optimization

Version, test, and deploy prompts with full lineage tracking. Automatically optimize prompts with state-of-the-art algorithms to improve performance.

AI Gateway

Unified OpenAI-compatible API gateway for all LLM providers. Route requests, manage rate limits, handle fallbacks, and control costs from one interface.

Experiment Tracking & Model Registry

Track ML experiments, evaluate models, register production models, and deploy them with MLflow's model lifecycle tools.

Technical capabilities

OpenTelemetry-Based Tracing

Tracing is built on OpenTelemetry, providing vendor-neutral instrumentation and compatibility with any LLM provider and agent framework.

Agent Server

FastAPI-based hosting solution with automatic request validation, streaming support, and built-in tracing for deploying agents to production.

Must watch videos

MLflow MCP Registry: Register, Version, and Manage Custom MCP Servers

19:56

Tutorial19:56

MLflow MCP Registry: Register, Version, and Manage Custom MCP Servers

Learn how to write, register, and manage custom and third-party MCP servers with the MLflow MCP registry. The registry is a central place to register, version, discover, and alias MCP servers so tools can be tracked in an MLflow tracking server. It works similarly to a prompt registry. This tutorial covers: 🔹 Three registry entities: MLflow Server (a logical server with a canonical namespace slug, for example io.github...), MCP Server 🔹 Version (snapshot and version a set of tools as they move from staging to development to production), and Access Point (how the MCP client connects, via standard IO locally over IPC or stream HTTP with JSON-RPC over HTTP) 🔹 Lifecycle stages: Draft, Active, Deprecated, and Deleted. An active server cannot be deleted until it has been deprecated first 🔹 Building a custom MCP server: a Python FastMCP example called "order analytics," an in-memory fake database, with tools created by decorating functions with @mcp.tool 🔹 Registering servers: JSON config for the custom server and a third-party Wikipedia server (deep-wiki), then querying them with the MCP client API and with natural language in Claude 🔹 Versioning in the UI: bump a version, apply a production alias, and keep multiple active versions (staging and production) side by side 🔹 Cleanup: a helper that kills active subprocesses and unregisters entities so the workflow stays idempotent Notebook: https://github.com/dmatrix/mlflow-genai-tutorials/blob/main/mlflow_mcp_registry.ipynb Speaker: Jules Damji, Developer Advocate, Databricks 00:00 - Introduction to the MLflow MCP Registry 01:06 - MLflow MCP Registry Architecture Overview 02:08 - Understanding the Three Registry Entities 04:58 - Environment Setup & Prerequisite Packages 06:02 - Building a Custom MCP Server in Python 08:33 - Registering a Custom Server in the MLflow UI 10:19 - Querying the Custom Server Using the Client API 12:58 - Registering and Accessing an External Server (Wikipedia) 14:24 - Querying the External Server via Natural Language (Claude) 16:44 - Managing Versions, Lifecycle Stages, and Workspace Cleanup #MLflow #MCP #Databricks #GenAI #Python #Claude #ModelRegistry

MLflow MCP Server Tutorial: Query Traces with Python and Natural Language

16:45

Tutorial16:45

MLflow MCP Server Tutorial: Query Traces with Python and Natural Language

Learn how to use the MLflow MCP (Model Context Protocol) server to talk to your tracking data from code or from an AI assistant. MCP lets clients and AI agents call external services as tools. The MLflow MCP server can run locally or alongside a tracking server, and it exposes about 20 to 30 function tools for both read and write operations: analyze latencies, diagnose errors, log traces with annotations, and delete data. This tutorial covers two ways to interact with the server: 🔹 Programmatic: connect asynchronously with the Python fastmcp client and call specific tools, including filtering by genai or traditional ml categories 🔹 Natural language: configure Claude, VS Code, or Cursor with a JSON file and query tracking data in plain English Live demos include: 🔹 Seeding traces: simulated experiments with high-latency tasks, flaky tools/exceptions, RAG chains, and chat token usage 🔹 Querying and filtering: error-status traces, executions longer than one second, production tags, and token count breakdowns 🔹 Writing back and cleanup: save trace annotations to the server, retrieve them, then delete traces so the data stays idempotent 🔹 Notebook: https://github.com/dmatrix/mlflow-genai-tutorials/blob/main/mlflow_mcp_server.ipynb Speaker: Jules Damji, Developer Advocate (Databricks) 00:00 – Introduction to the MCP 00:39 – Architecture and Capabilities of the MLflow MCP Server 02:35 – Setting up the Local Tracking Server and Environment 03:26 – Seeding Fake Experimental Data and Trace Types 05:05 – Programmatic Discovery: Categorizing Tools via FastMCP 07:29 – Two Approaches to Interacting with the MCP Server 08:13 – Programmatic Querying: Filtering Errors and Measuring Latency 09:48 – Writing Data Back: Logging Trace Annotations and Hierarchy 12:36 – Natural Language Querying 14:02 – Live Claude Demo MLflow, MCP, GenAI, Machine Learning, Python, Cursor, Claude, VSCode, Observability

How to Trace and Debug AI Coding Agents with Omnigent and MLflow

8:19

Tutorial8:19

How to Trace and Debug AI Coding Agents with Omnigent and MLflow

Managing multi-agent workflows across different coding harnesses can lead to fragmented context, lost guardrails, and zero visibility. In this video, discover how to solve these problems by integrating the open-source meta-harness, Omnigent, with MLflow. Learn how Omnigent unifies context and policies across diverse coding tools while delegating tasks to agents like CodeX and Claude. See step-by-step how to launch an OpenTelemetry environment, run a local multi-agent workflow, and use MLflow Traces to track exact tool calls, measure latency, and monitor token usage. What you will learn: 🔹 The challenges of switching between multiple AI coding harnesses. 🔹 How Omnigent provides a unified interface and shared context across agent boundaries. 🔹 Setting up your environment and tracking live multi-agent execution. 🔹 Using the MLflow UI to analyze traces, inspect tool chains, and debug failures. Links & Resources: 🔗 MLflow Documentation: https://mlflow.org/docs/latest/index.html 🔗 Omnigent Repo: https://github.com/databricks/omnigent 🔗 MLflow LLM Tracking & Tracing Guide: https://mlflow.org/docs/latest/llms/tracing/index.html

Who uses it

Real use cases, no hype

LLM & Agent Tracing

Debug multi-step agent runs with full OpenTelemetry spans across any provider.

LLM Evaluation & Quality Gates

Score outputs with 50+ built-in judges, track regressions, and block bad releases.

Prompt Versioning & Optimization

Version prompts with lineage and auto-tune them against your datasets.

AI Gateway & Cost Governance

Route across LLM providers, enforce rate limits, and cap spend behind one API.

Classical ML Experiment Tracking

Log params, metrics, and artifacts for PyTorch, sklearn, XGBoost and more.

Model Registry & Deployment

Version models, promote stages, and serve them with consistent packaging.

Production AI Monitoring

Continuously evaluate quality, detect drift, and track costs on live traffic.

What's great

100% open source, Apache 2.0, no license fees
One-command setup with `uvx mlflow server`
100+ integrations: LangChain, OpenAI, PyTorch, etc.
Single platform for classical ML and LLM/agent workflows

Technical strengths

OpenTelemetry-native tracing for any provider
Prompt registry with lineage and automated optimization
AI Gateway with rate limits, fallbacks, cost control

Where it falls short

UI gets sluggish with thousands of runs
Self-hosting at scale means managing Postgres + S3 yourself
Docs split across 'LLMs' and 'ML' sections is confusing
LLM-as-judge evals burn API tokens fast

Technical limitations

Default `./mlruns` file store is not production-safe
Prompt optimization and AI insights are newer, less battle-tested
Some autolog integrations break on framework version bumps

For technical folks

The deep-dive specs

License

Apache 2.0

Monthly Downloads

30M+

GitHub Stars

20K+

Contributors

900+

Supported Languages

Python, TypeScript/JavaScript, Java, R

Integrations

100+ tools

Tracing Protocol

OpenTelemetry

Agent Server

FastAPI-based

FAQ

Includes technical Q&A

An open source AI engineering platform covering LLM/agent observability, evaluation, prompt management, and the classical ML lifecycle (tracking, registry, deployment).