Library of the Week — Marvin

Marvin — the AI engineering toolkit that adds typed, LLM-powered functions to your Python code

GitHub · Language: Python · License: Apache 2.0

What it does

Marvin lets you define AI-powered functions using plain Python type annotations — no prompt templates, no chains, no boilerplate. It’s built for developers who want LLM capabilities (classification, extraction, generation, transformation) that feel native to Python rather than bolted on.

Why it stands out

• Decorator-first API: @marvin.fn turns a typed Python function signature into a full LLM call — the docstring is the prompt, the return type is the output schema, enforced automatically
• Built-in task primitives: marvin.extract(), marvin.classify(), marvin.cast(), and marvin.generate() cover the most common NLP tasks with zero configuration
• Pydantic-native output: return types are validated Pydantic models by default, so you get structured, type-safe results without wiring up a separate validation layer
• Model-agnostic: works across OpenAI, Anthropic, and other providers through a clean settings interface — swap models without touching task code

Quick start

import marvin
from pydantic import BaseModel

class Sentiment(BaseModel):
    label: str
    confidence: float

# Extract structured data from unstructured text
results = marvin.extract(
    "The new API is fast but the docs are confusing and pricing feels steep.",
    target=Sentiment,
    instructions="Extract each distinct sentiment expressed."
)

for r in results:
    print(r.label, r.confidence)

When to use it

• You need quick, reliable structured extraction or classification inside an existing Python codebase and don’t want to build a prompt pipeline from scratch
• You’re prototyping product features that need LLM capabilities (tagging, routing, parsing) and want something that composes naturally with the rest of your app logic
• Your team is Python-first and finds chain-based frameworks like LangChain too abstraction-heavy for simple tasks

When to skip it

• You’re building complex multi-agent workflows with intricate tool-use graphs — Marvin doesn’t have native agent orchestration; something like Smolagents or a custom loop fits better there
• You need fine-grained control over prompt construction, few-shot examples, or retrieval pipelines; Marvin intentionally hides that layer, which becomes a liability at scale

The verdict

Marvin occupies a smart niche: it’s not trying to be a full agent framework, it’s trying to make individual LLM-powered operations feel like regular Python. For teams adding AI capabilities to existing applications — rather than building AI-first systems from scratch — that’s exactly the right tradeoff. If you’ve ever written a 40-line extraction function using raw API calls and thought “there has to be a better way,” Marvin is the better way.