mlflow--mlflow/mlflow/agent/setup/templates/python.md

1. Install MLflow

Detect this project's Python package manager and add mlflow as a dependency if it is not already declared:

• uv (look for uv.lock or [tool.uv] in pyproject.toml) -> uv add mlflow
• poetry (look for poetry.lock) -> poetry add mlflow
• pip / plain requirements.txt -> append mlflow and pip install mlflow

Skip this step if mlflow is already a declared dependency.

{{ server_setup }}### 2. Configure tracking URI

Configure MLflow to log to {{ tracking_uri }}. Pick whichever of these fits the project's conventions:

• Set MLFLOW_TRACKING_URI={{ tracking_uri }} in the project's env file (.env, .env.example, etc.).
• Call mlflow.set_tracking_uri("{{ tracking_uri }}") once during application startup, before any mlflow.* calls.

Don't do both. If the project already sets a tracking URI, leave it alone and note the existing value in the final summary.

3. Instrument with mlflow.autolog

Consult the instrumenting-with-mlflow-tracing skill in {{ skills_dir }}/ for the supported libraries and per-integration setup. That skill is the source of truth for what mlflow.autolog() covers.

For most applications, mlflow.autolog() is the recommended entry point:

import mlflow

mlflow.set_tracking_uri("{{ tracking_uri }}")
mlflow.autolog()

Wire this into the application's entry point(s):

• Find the main entry (e.g. main.py, app.py, __main__.py, FastAPI lifespan / Depends, Django app config ready hook, Lambda handler init).
• Call mlflow.autolog() once, before any LLM clients are created.
• Do not add it to library modules or tests.

For library-specific instrumentation (LangChain, LangGraph, OpenAI, Anthropic, LlamaIndex, DSPy, etc.), many libraries have a dedicated mlflow.<library>.autolog() flavor. The skill above lists them.