Marvin

Marvin is a Python framework for two related jobs: getting structured, typed output from large language models (LLMs), and orchestrating agentic workflows around those calls. The two layers sit in one package, so a single typed extraction and a multi-step agent run share the same API.

The structured-output layer is a handful of top-level functions. marvin.extract pulls native types out of unstructured text, marvin.cast converts input into a target type such as a TypedDict, marvin.classify maps input onto the members of an enum, marvin.generate produces a requested number of structured objects from a description, and marvin.summarize condenses text. Each takes the Python type as the target and returns a validated value of it.

The agentic layer is built from tasks, agents, and threads. A Task carries instructions, a result_type, optional tools, and context, and returns a type-safe result. An Agent is a portable configuration of model, instructions, and tools that can be assigned to any task and reused across workflows. A Thread is a context manager that shares message history and context across several runs. marvin.plan breaks an objective into dependent tasks, which marvin.run_tasks then executes.

Tool use is plain Python: a function passed in the tools list becomes callable by the agent, and each call and its output print as a discrete step during the run. All Marvin functions have thread management built in, so they compose into chains of tasks that share context.

Version 3.0 combines the earlier Marvin interface with the agentic engine from ControlFlow, which it supersedes, and renames ControlFlow’s Flow to Thread. Model calls go through Pydantic AI, so the full range of providers Pydantic AI supports is available, with OpenAI as the default. Thread and message history is stored in SQLite, and no database migrations are available yet, so data may need resetting between updates.