How much does MLflow cost?

Free and open source under the Apache 2.0 license, with no seat fees, usage limits or paid tiers from the project. Costs are the infrastructure a team runs the tracking server on and the model calls made by LLM judges, issue detection and AI Gateway traffic on the team's own provider keys. Managed MLflow is sold separately by Databricks, AWS SageMaker, Nebius and others at their own prices.

What are the best alternatives to MLflow?

Based on documented capability coverage across the Agentic Index 14 point taxonomy, the closest alternatives to MLflow among agent infrastructure platforms are Arize AI, Confident AI, HoneyHive, Langfuse. Each is graded on the same public evidence, and no vendor pays for placement.

Does MLflow support MCP and tool calling?

Full under the Agentic Index verification standard. MLflow publishes public evidence of Model Context Protocol (MCP) support, tool calling, and API or SDK extensibility for connecting agents to external tools. Evidence basis: A REST API on the tracking server sits under SDKs for Python, TypeScript and JavaScript, Java and R, and OpenTelemetry ingestion accepts traces from any language. The MLflow MCP server, experimental and launched locally over stdio against t. Graded from public evidence only against the Agentic Index 14 point capability taxonomy. No vendor pays for placement.

Does MLflow have SOC 2, SSO and role based access control?

Full under the Agentic Index verification standard. MLflow publishes public evidence of enterprise security and identity controls, covering a compliance posture such as SOC 2 or ISO 27001, single sign on (SSO or SAML), and role based access control (RBAC). Evidence basis: Built in authentication adds role based access control with read, use, edit and manage levels plus an explicit deny, across experiments, prompts, registered models, scorers and gateway endpoints, with platform admins, workspace managers and. Graded from public evidence only against the Agentic Index 14 point capability taxonomy. No vendor pays for placement.

Can MLflow be self hosted or run in a private VPC?

Full under the Agentic Index verification standard. MLflow publishes public evidence of self hosted or private VPC deployment, along with data residency options for controlling where agent data is processed and stored. Evidence basis: Everything runs where the customer puts it. The tracking server starts locally with one command on SQLite, moves to PostgreSQL or MySQL with an artifact store for production, ships as a Docker Compose project and a Helm chart for Kubernetes. Graded from public evidence only against the Agentic Index 14 point capability taxonomy. No vendor pays for placement.

Back to vendors

MLflow

Q: Who is MLflow for?

AI engineering teams building agents on any framework, machine learning teams already running MLflow, platform teams that self host AI tooling, organizations that keep agent telemetry inside their own network

Also known as: MLflow Tracing, MLflow AI Gateway, MLflow Prompt Registry, MLflow GenAI, mlflow.org

Visit site

Agent infrastructureindependentVerified 2026-11-09

Entry priceFree and open source under Apache 2.0, with no seat fees or usage limits. Teams pay only for the infrastructure they run it on and the model calls its judges and gateway make on their own keys.Full pricing detail

MLflow is an open source AI engineering platform, hosted by the Linux Foundation, for teams building agents and LLM applications. Tracing on OpenTelemetry records every model call, tool call and retrieval with cost and latency, more than 70 built in LLM judges and custom scorers test agents before release and on live traffic, a prompt registry versions and optimizes prompts, and an AI Gateway routes calls across providers with fallbacks, guardrails and budgets. It is free under Apache 2.0 and runs on the team's own infrastructure.

MLflow is an open source AI engineering platform for agents, LLM applications and machine learning models, released under the Apache 2.0 license and hosted by the Linux Foundation as MLflow Project, a Series of LF Projects, LLC. Databricks holds the original copyright on the code, and the project says no single vendor controls it. It counts more than 30 million package downloads a month, more than 900 contributors and more than 100 integrated tools. Beside its agent tooling it keeps the machine learning features it started with: experiment tracking, model evaluation, model packaging and a model registry.

For agents, MLflow covers the loop from first trace to production monitoring. Tracing built on OpenTelemetry instruments more than 40 frameworks and providers with one line, recording each LLM call, tool call and retrieval with tokens, cost and latency, grouped by user and session. Evaluation runs an agent against a dataset with more than 70 built in LLM judges for correctness, safety, groundedness and tool call quality, plus code scorers and judges written in plain language, and regression tests in pytest gate changes in CI.

The same judges run automatically on a sampled share of production traces inside the server, issue detection groups recurring failures such as an invented policy or a repeated tool call, and review queues send traces to domain experts. The Prompt Registry versions prompts with aliases and lineage and optimizes them against scores. AI Gateway puts one OpenAI compatible endpoint in front of more than a dozen providers with traffic splits, fallbacks, guardrails, budgets and usage tracking. An Agent Server hosts an agent behind a REST endpoint, and an MCP server lets coding assistants search traces and log feedback.

MLflow costs nothing to use, with no seat or usage fees. Spend is the infrastructure a team runs it on and the model calls its judges and gateway make on the team's own keys. Managed MLflow is sold by Databricks, AWS SageMaker, Nebius and others on their own terms.

Self hosting runs from one command on SQLite up to PostgreSQL with Docker Compose or a Helm chart on Kubernetes, on any cloud or on premises, so trace data stays where the team puts it. Built in authentication adds role based access control with workspaces, single sign on runs through a community OIDC plugin, provider keys are stored encrypted, and vulnerabilities go through coordinated disclosure. As a project rather than a hosted service, it holds no SOC 2 or ISO attestation.

It fits engineering teams that want tracing, evaluation, prompt management and a model gateway in one open source system they host, machine learning teams already on MLflow who are adding agents, and organizations that keep agent telemetry inside their own network. It fits less well for a team that wants a hosted service with a contract and support terms from the project itself, one that needs an agent runtime with memory, retrieval or workflow orchestration, which MLflow leaves to the agent framework, or one that needs a person to approve an agent's actions before they run.

Vendor details

Canonical URL

https://mlflow.org

Subcategory

Open source tracing, evaluation, prompt management and model gateway for the agents and LLM applications a team builds

Funding status

Open source project under the Apache 2.0 license, hosted by the Linux Foundation as MLflow Project, a Series of LF Projects, LLC. Databricks holds the original copyright on the code, and the project says no single vendor controls it. More than 900 contributors and more than 20,000 GitHub stars.

Company status

independent

Use cases & customers

Primary use cases

tracing agent runs, tool calls, tokens and cost with OpenTelemetryevaluating agents with LLM judges before release and on production trafficgating agent changes with regression tests in CIversioning and optimizing promptsrouting model calls across providers with fallbacks, guardrails and budgets

Target customers

AI engineering teams building agents on any frameworkmachine learning teams already running MLflowplatform teams that self host AI toolingorganizations that keep agent telemetry inside their own network

Deployment options

Self hosted tracking server, local or on any cloudDocker ComposeKubernetes with the Helm chartOn premises

Integrations

Automatic tracing covers more than 40 libraries, including LangChain, LangGraph, the OpenAI Agents SDK, Google ADK, CrewAI, DSPy, Pydantic AI, AutoGen, LlamaIndex, Haystack, the Claude Agent SDK and the Vercel AI SDK, plus OpenTelemetry from any language. AI Gateway connects to OpenAI, Anthropic, Google Gemini, Azure OpenAI, Amazon Bedrock, Vertex AI, Cohere, Mistral, Groq, Together AI, Fireworks AI, Ollama and Portkey, with LiteLLM for others. Third party scorers come from DeepEval, RAGAS, Phoenix, TruLens and Guardrails AI. The MCP server works with Claude Code, Cursor and other assistants, and SDKs cover Python, TypeScript, Java and R.

In practice

A support team running a LangGraph agent adds one autolog line pointed at its self hosted MLflow server. A groundedness judge scores 10% of production traces, and when scores drop after a retrieval change, an engineer opens a low scoring trace and finds the retriever returning last year's refund policy.

An insurance platform team sends every agent's model calls through MLflow AI Gateway, splitting traffic 80 percent and 20 percent between two models to compare them, with a cheaper fallback for rate limits, a guardrail that redacts personal data and a monthly budget that rejects traffic past $5,000.

Before a prompt change merges, the pull request runs MLflow regression tests in pytest where guideline judges assert that the agent still refuses to reveal its system prompt and answers in one sentence. A failing judge blocks the merge, and the run stays in MLflow for the team to inspect.

Sources & related URLs

Related / legacy domains

https://mlflow.org/docs/latest/genai/ https://github.com/mlflow/mlflow https://mlflow.org/faq/

Research sources

Agentic Index coverage score

9.0 / 14 capabilities · 64%

Integrations & Tool Calling	Partial
Agents built on LangChain, LangGraph, the OpenAI Agents SDK, Google ADK, CrewAI, DSPy, Pydantic AI and the rest of more than 40 supported libraries send traces into MLflow with one line of setup, and OpenTelemetry carries the rest. The MCP Registry, still experimental, catalogs MCP servers with versions, aliases and access endpoints that tell an agent where to connect, but it stores no credentials to outside systems and calls no tools itself. Actions an agent takes in other software run inside the customer's own framework code. Sourcemlflow.org/docs/latest/genai/mcp-registryread 2026-11-09
Workflow Orchestration	Not documented
Agent logic lives in whatever framework the team picks, and MLflow sits beside it rather than running the steps. The Agent Server is a FastAPI host that exposes one agent function at an invocations endpoint with request validation, streaming and automatic tracing, and the ResponsesAgent interface packages tool calls and outputs from several agents when the team's own code coordinates them. There is no workflow engine, graph builder or step scheduler, and no retry or branching logic for agent steps. Sourcemlflow.org/docs/latest/genai/serving/agent-serverread 2026-11-09
Knowledge Grounding & RAG	Not documented
Retrieval is something MLflow measures rather than provides. RAG judges score whether retrieved documents are relevant to the request, whether the answer is grounded in them and whether they hold everything needed, and traces from LlamaIndex, LangChain and Haystack show each retrieval step with its inputs and results. There is no vector index, document store or knowledge base that an agent queries to ground its answers; that layer sits in the customer's own stack. Sourcemlflow.org/docs/latest/genai/eval-monitor/scorers/llm-judge/ragread 2026-11-09
Human Oversight & Guardrails	Partial
Guardrails on AI Gateway endpoints run an LLM judge before a request reaches the model and after the response returns, checking for unsafe content, personal data and custom policies, then block the call with an HTTP 400 or redact the flagged text; streamed responses skip the post response check. Budget policies can reject further gateway traffic once spend passes a threshold, and prompt aliases such as production decide which version an app loads, behind edit permissions. Nothing pauses an agent's action for a person to approve. Sourcemlflow.org/docs/latest/genai/governance/ai-gateway/guardrailsread 2026-11-09
Security, Identity & Governance	Full
Built in authentication adds role based access control with read, use, edit and manage levels plus an explicit deny, across experiments, prompts, registered models, scorers and gateway endpoints, with platform admins, workspace managers and isolated workspaces. Single sign on comes through a community maintained OIDC plugin with group mapping for Okta, Google, AWS Cognito and Entra ID, or through a reverse proxy. The server guards against DNS rebinding, CORS abuse and clickjacking, blocks SSRF on gateway URLs and keeps provider keys encrypted. As software the customer runs, the project holds no SOC 2 or ISO 27001 attestation. Sourcemlflow.org/docs/latest/self-hosting/security/role-based-access-controlread 2026-11-09
Observability & Auditability	Full
Tracing built on OpenTelemetry records every LLM call, tool call, retrieval and agent step as spans with inputs, outputs, latency, token counts and estimated cost in USD, rolled up per trace and tagged by user, session, version and environment. Traces can be searched, filtered, charted on an overview dashboard, exported over OTLP to other backends, archived or deleted, and personal data can be masked before a span leaves the application. AI Gateway also logs requests and responses for audit trails. Sourcemlflow.org/docs/latest/genai/tracingread 2026-11-09
Memory & State Persistence	Not documented
Session and user IDs attach to traces so a team can filter and group a multi turn conversation after the fact, and the ResponsesAgent interface passes chat history through on each request. Neither is a store the agent reads back as context in a later session. The MemAlign optimizer keeps semantic and episodic memory, but it tunes LLM judges from reviewer feedback rather than remembering anything for the customer's agent, and it is experimental. Sourcemlflow.org/docs/latest/genai/tracing/track-users-sessionsread 2026-11-09
Deployment & Data Residency	Full
Everything runs where the customer puts it. The tracking server starts locally with one command on SQLite, moves to PostgreSQL or MySQL with an artifact store for production, ships as a Docker Compose project and a Helm chart for Kubernetes, and runs on any cloud or on premises, so traces, prompts and provider keys stay on infrastructure the customer controls. An experimental option runs code based scorers inside job containers with no network access. Sourcemlflow.org/docs/latest/self-hostingread 2026-11-09
Prebuilt Agents / Templates / Packs	Full
Ready made judges ship for response quality (relevance, correctness, completeness, fluency, safety, guidelines), retrieval (relevance, groundedness, sufficiency), tool calls (correctness, efficiency) and multi turn sessions, more than 70 in all, and third party scorer packs from DeepEval, RAGAS, Phoenix, TruLens and Guardrails AI plug in beside them. The MLflow AI Assistant turns a coding agent such as Claude Code into a setup helper that adds tracing and evaluation to a codebase, and an experimental Skill Registry versions agent skills. Sourcemlflow.org/docs/latest/genai/eval-monitor/scorers/llm-judge/predefinedread 2026-11-09
Triggers & Channel Coverage	Full
Automatic evaluation starts on its own as traces land: LLM judges run inside the tracking server on a chosen share of incoming traces or completed sessions, from 0 to 100 percent, filtered by trace search syntax, with no person starting each run. Issue detection can be scheduled daily or weekly to analyze new traces. Webhooks, still experimental, send HTTP POST calls on model and prompt registry events and on AI Gateway budget breaches, and deployed agents take requests over a REST endpoint. Sourcemlflow.org/docs/latest/genai/eval-monitor/automatic-evaluationsread 2026-11-09
Model Flexibility & Routing	Full
Teams choose the models themselves. AI Gateway endpoints connect to OpenAI, Anthropic, Google Gemini, Azure OpenAI, Amazon Bedrock, Vertex AI, Cohere, Mistral, Groq, Together AI, Fireworks AI, Ollama for local models and Portkey, with LiteLLM extending the list, and an admin sets weighted traffic splits across models for A/B tests and ordered fallbacks for outages and rate limits. LLM judges and issue detection run on whichever provider, model or gateway endpoint the team selects. Sourcemlflow.org/docs/latest/genai/governance/ai-gateway/endpoints/model-providersread 2026-11-09
APIs / SDKs / MCP Extensibility	Full
A REST API on the tracking server sits under SDKs for Python, TypeScript and JavaScript, Java and R, and OpenTelemetry ingestion accepts traces from any language. The MLflow MCP server, experimental and launched locally over stdio against the team's tracking URI, gives Claude Code, Cursor and other assistants ten tools to search and fetch traces, log and update feedback and expectations, set or remove trace tags and delete traces. Prompts load into applications through the SDK by name and alias. Sourcemlflow.org/docs/latest/genai/mcpread 2026-11-09
Testing, Debugging & Optimization	Full
Each change to an agent can be tested before release: an evaluation runs it against a dataset with LLM judges and code scorers, and an experimental comparison applies paired significance tests with confidence intervals to show whether one run truly beats another. Regression tests written as pytest functions assert that judges pass and gate pull requests in CI. Conversation simulation generates multi turn test cases, review queues send traces to people whose answers feed evaluation, and judge alignment and prompt optimization tune against the scores. Sourcemlflow.org/docs/latest/genai/eval-monitor/regression-testingread 2026-11-09
Browser / Computer-use	Not documented
Agents traced or evaluated by MLflow may drive a browser in their own code, but MLflow itself controls no browser, desktop or remote computer session. The coding agent setup wizard and AI Assistant run in the developer's own terminal through Claude Code, Codex or OpenCode, and the tracing integrations cover model providers and agent frameworks rather than an interface automation layer the project operates. Sourcemlflow.org/docs/latest/genairead 2026-11-09

The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded

Pricing

Free and open source under Apache 2.0, with no seat fees or usage limits. Teams pay only for the infrastructure they run it on and the model calls its judges and gateway make on their own keys.

No subscription or metering from the project. Costs are the team's own hosting and the model provider usage behind LLM judges, issue detection and AI Gateway traffic.

Free tier

Included quota

Everything: tracing, evaluation, automatic evaluation, issue detection, review queues, the Prompt Registry, AI Gateway, the MCP server and the Agent Server, with no seat, trace or usage limits.

What is public

MLflow is open source under the Apache 2.0 license and free for any purpose, including commercial use, with no licensing fees, seat fees or usage limits. The project names managed MLflow from Databricks, AWS SageMaker, Nebius and others for teams that do not want to run it themselves.

Billing mechanics

The project bills nothing. Model calls from judges, issue detection and gateway traffic are billed by each provider on the team's own key, and hosting costs follow the team's own cloud or data center.

Cost watchouts

Automatic evaluation and issue detection call a model for every sampled trace or scheduled run, billed on the team's provider key. Running the tracking server in production takes a SQL database, artifact storage and operations time.

Variable cost rationale

MLflow itself is free and meters nothing. The variable parts are the team's own: the servers, database and artifact storage the tracking server runs on, and model provider usage from LLM judges, automatic evaluation on production traces, issue detection and AI Gateway traffic. Trace sampling rates and gateway budget policies limit how far those grow.

Additional watchouts

No support contract or SLA comes from the project. Several agent features, including the MCP server, MCP Registry, review queues and webhooks, are still experimental and may change between releases.

Overage / add-ons

None. The project meters nothing. Model usage is billed by the provider behind each key, and AI Gateway budget policies can alert on or reject traffic past a set dollar threshold.

Sales call required

No, self serve available

Free / trial

Free and open source under Apache 2.0, self hosted, with every feature included.

Lowest paid plan

No paid plan. The project sells nothing, and every feature is free under Apache 2.0.

Commercial notes

Hosted by the Linux Foundation as MLflow Project, a Series of LF Projects, LLC, with Databricks holding the original copyright on the code. The site counts more than 30 million downloads a month and more than 900 contributors.

Key ambiguities

There is no MLflow price. Managed MLflow from Databricks, AWS SageMaker, Nebius and other providers is priced by each provider and is a separate purchase.

Cancellation / refund

No contract. The software is used under the Apache 2.0 license.

Support SLA / resale

Community support through GitHub issues, Slack, a mailing list and office hours. The project offers no paid support or SLA; managed providers set their own.

Missing data

No vendor price exists. Costs depend on hosting choices and model usage, which differ for every team.

Agentic Index verified 2026-11-09

Official pricing page Docs

Related vendors

Alternatives to MLflow

The closest documented capability profiles to MLflow among agent infrastructure platforms tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.

Arize AI9.0 / 14Matches MLflow across all 14 documented capabilities
Confident AI8.5 / 14A lighter documented profile than MLflow
HoneyHive8.5 / 14A lighter documented profile than MLflow
Langfuse8.5 / 14A lighter documented profile than MLflow
F5 AI Guardrails9.0 / 14Fuller documented coverage on Human Oversight & Guardrails
Freeplay8.0 / 14A lighter documented profile than MLflow

Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded

Compare MLflow with other vendors →Compare all agent infrastructure platforms →See where MLflow ranks →

Data confidence: high