Marvin

Marvin is Prefect's Python framework for producing structured outputs and building agentic AI workflows. Version 3.0 fused Marvin 2.x's structured-output developer experience (cast, classify, extract, generate, summarize) with the agentic engine of ControlFlow — which Marvin 3.0 supersedes — giving you one-liner task runs (marvin.run), specialized Agents, discrete observable Tasks, and Threads that compose tasks with shared context and history.

All LLM interactions go through Pydantic AI, so Marvin natively supports every model provider Pydantic AI supports, with OpenAI as the default. Thread/message history is stored in SQLite (no migrations yet during development, per the provider's own upgrade notes).

1. Install: uv add marvin or pip install marvin
2. Configure your LLM provider — OpenAI by default: export OPENAI_API_KEY=<your-key> (any Pydantic AI-supported model/provider works natively)
3. Try the structured-output utilities: marvin.extract(...), marvin.cast(...), marvin.classify(...), marvin.generate(...)
4. Build agentic flows: marvin.run("Write a short poem..."), or define marvin.Agent / marvin.Task explicitly and compose them inside a marvin.Thread for shared context across tasks