Library of the Week — Magentic

Magentic — turn Python functions into LLM-powered calls with a single decorator

GitHub · Language: Python · License: MIT

What it does

Magentic lets you replace a Python function body with an LLM call by decorating it with @prompt, automatically handling serialization, structured output via Pydantic, and async support. It’s aimed at developers who want LLM integration that feels like ordinary Python — no chains, no graph, no boilerplate.

Why it stands out

• Decorator-first design keeps LLM logic colocated with the function signature, making intent obvious and diffs clean — you never reach for a prompt template buried in a config file
• Native structured output infers the return type from the annotation and uses function-calling or JSON mode under the hood — return a list[Item], get a list[Item], not a string to parse
• Multi-backend support spans OpenAI-compatible endpoints, Anthropic, and local models via LiteLLM integration, so swapping providers is a one-line change
• Async and streaming are first-class: decorate with @prompt on an async def and it just works; streaming returns an AsyncIterator without extra setup

Quick start

from pydantic import BaseModel
from magentic import prompt

class Flashcard(BaseModel):
    question: str
    answer: str

@prompt("Generate a flashcard for the topic: {topic}")
def make_flashcard(topic: str) -> Flashcard: ...

card = make_flashcard("transformer attention mechanisms")
print(card.question)
print(card.answer)

When to use it

• You want structured LLM outputs inside an existing Python codebase without adopting a full framework like LangChain or LlamaIndex
• Your team thinks in functions and type annotations — Magentic fits that mental model instead of fighting it
• You need to iterate quickly on prompt logic while keeping the surrounding code testable and statically analyzable

When to skip it

• You need complex multi-step orchestration, memory, or tool-use loops — at that point a dedicated agent framework (LangGraph, Agno) gives you more control
• Your production stack is TypeScript-first; Magentic is Python-only with no official JS port

The verdict

Magentic is the cleanest “just make this function call an LLM” library in the Python ecosystem — it earns its keep precisely because it knows what it isn’t. If you’re tired of frameworks that make simple things complicated, start here and only reach for something heavier when you genuinely need it.