MLflow for LLMs and agents: pricing, licence and features
by the MLflow project (Linux Foundation)
Checked 2026-10-20 on MLflow's own docs and GitHub repository. Anything we could not confirm says “not verified”. We don't score or rank tools: how we review.
Short answer: MLflow is a free, Apache-2.0 platform that now covers LLM and agent work as well as classical ML: OpenTelemetry-compatible tracing, LLM-as-a-judge evaluation, a prompt registry and an AI Gateway. It suits teams that already use MLflow for model tracking, or want one open-source system for both. There is no MLflow-run hosted plan; you self-host it or buy it managed from a cloud provider.
Pricing and deployment
| What the source says | Source | |
|---|---|---|
| Licence | Apache-2.0 | GitHub repo |
| Governance | Linux Foundation project, described as vendor-neutral | MLflow docs |
| Self-hosting cost | No licence fee; you pay for your own server, database, artifact store and judge-model calls | Self-hosting |
| Managed options | Databricks, AWS SageMaker, Azure Machine Learning, Nebius, GCP (GKE); prices set by each provider (not verified) | Self-hosting: cloud services |
| Quickest start | pip install mlflow, then mlflow server (SQLite backend by default since 3.7.0) | Self-hosting |
| Production setups | Docker Compose (MLflow + PostgreSQL + RustFS) or the official Helm chart; PostgreSQL plus S3, GCS or Azure Blob for scale | Self-hosting |
| Access control | Basic HTTP auth, SSO and custom auth plugins; opt-in workspaces with workspace-level permissions | Self-hosting |
LLM and agent features
| Feature | MLflow | Source |
|---|---|---|
| Tracing | Captures prompts, retrievals, tool calls and LLM responses; OpenTelemetry-compatible with GenAI semantic conventions | MLflow docs |
| LLM-as-a-judge | Pre-built judges (e.g. hallucination, relevance) and custom judges; custom metrics | MLflow docs |
| Prompt management | Version, compare and reuse prompt templates with lineage; automated prompt optimisation | MLflow docs |
| AI Gateway | Manage model access and costs across providers | MLflow docs |
| Integrations | The docs cite 100+ framework and provider integrations | MLflow docs |
| Classical ML | Experiment tracking, registered models, deployment | Self-hosting |
| Human annotation queues | not verified | — |
| SOC 2 / HIPAA | Depends on the managed provider; not applicable to the open-source software itself | — |
Who should pick MLflow
- Pick it if you already run MLflow (or Databricks, SageMaker or Azure ML) for model work and want LLM tracing and evals in the same place.
- Pick it if you need an Apache-2.0 licence and a vendor-neutral foundation project rather than a single-company product.
- Look elsewhere if you want a hosted plan with a published monthly price from the tool's own maker: Langfuse, Comet Opik and Braintrust publish theirs.
- Moving off Helicone? MLflow is one of the options with a gateway; see the Helicone migration guide.
Compare
FAQ
Is MLflow free?
The software is free: MLflow is Apache-2.0 open source and a Linux Foundation project. Running it yourself costs whatever the server, database, artifact storage and any LLM-judge model calls cost you. Managed MLflow from Databricks, AWS SageMaker, Azure Machine Learning, Nebius or GCP is priced by that provider; we have not verified those rates.
Is MLflow only for classical machine learning?
No. MLflow started with experiment tracking and a model registry, and its docs now position it as an AI engineering platform for LLMs and agents: OpenTelemetry-compatible tracing, LLM-as-a-judge evaluation, prompt versioning and optimisation, and an AI Gateway.
Does MLflow support OpenTelemetry?
Yes. MLflow's docs say its tracing is fully compatible with OpenTelemetry and supports the GenAI semantic conventions, so traces are not tied to MLflow's own SDK.
How do I self-host MLflow?
For one person or a small team, `pip install mlflow` and `mlflow server` start a server with a SQLite backend (the default since 3.7.0). The repository also ships a Docker Compose setup (MLflow, PostgreSQL and RustFS) and an official Helm chart for Kubernetes. For scale, the docs suggest PostgreSQL plus cloud object storage for artifacts.
MLflow or Langfuse?
If your team already runs MLflow for model training and the registry, adding its GenAI features avoids a second system. If you want a dedicated LLM platform with a published hosted price and annotation queues on every plan, Langfuse is built for that. Our Langfuse vs MLflow page goes through it point by point.
Does MLflow have SSO?
The self-hosting docs list username/password (basic HTTP auth), SSO and custom authentication plugins. Workspaces with workspace-level permissions are opt-in and need a SQL database backend.