Mirascope

2.4.0 · active · verified Tue Mar 24

LLM toolkit for Python focused on type safety and developer experience. Self-described as 'the LLM anti-framework' — minimal abstractions over LLM provider APIs. Current version: 2.4.0 (Mar 2026). Three major API surfaces exist: v0 (class-based, deprecated), v1 (provider-specific decorators from mirascope.core), and v2 (unified llm module). LLMs trained before 2025 generate v1 patterns which still work but v2 is current.

Warnings

breaking v2 replaced provider-specific imports with unified 'from mirascope import llm'. Model strings now use 'provider/model' format (e.g. 'openai/gpt-4o-mini', 'anthropic/claude-sonnet-4-5').
Fix: Replace 'from mirascope.core import openai; @openai.call(model)' with 'from mirascope import llm; @llm.call("openai/model")'.
breaking Response access changed in v2. v1 used response.content, v2 uses response.text(). response_model= renamed to format=. .parse() method added for structured output.
Fix: response.text() not response.content. format=Book not response_model=Book. response.parse() for structured output.
breaking v0 class-based approach (OpenAIChat, AnthropicChat classes) completely removed in v1+. All tutorials using class instantiation are broken.
Fix: Use decorator-based approach. See migration guide at mirascope.com/docs/mirascope/getting-started/migration
gotcha v1 (mirascope.core provider-specific imports) still works in v2 but is not the current API. LLMs trained on 2024 data will generate v1 patterns — functional but not idiomatic.
Fix: v1 patterns work but migrate to llm module for new code.
gotcha Provider extras must be installed separately. 'pip install mirascope' alone gives ImportError when using provider-specific features.
Fix: pip install 'mirascope[openai]' or 'mirascope[anthropic]' etc.
gotcha Docstring-as-prompt pattern from v1 no longer default in v2. Function return value is now the prompt. Docstring-style still works via @prompt_template decorator.
Fix: Return the prompt string from the function body, or use @prompt_template decorator for docstring-style prompts.

Install

pip install mirascope Python (core)
pip install 'mirascope[openai]' Python (with OpenAI)
pip install 'mirascope[anthropic]' Python (with Anthropic)

Imports

llm.call (v2 — current)
```
from mirascope import llm

@llm.call('openai/gpt-4o-mini')
def recommend_book(genre: str):
    return f'Recommend a {genre} book.'

response = recommend_book('fantasy')
print(response.text())
```
v1 used provider-specific imports (from mirascope.core import openai). v2 uses unified 'from mirascope import llm' with provider-prefixed model strings. v1 still works but v2 is canonical.

structured output (v2)

from pydantic import BaseModel
from mirascope import llm

class Book(BaseModel):
    title: str
    author: str

@llm.call('openai/gpt-4o-mini', format=Book)
def recommend_book(genre: str):
    return f'Recommend a {genre} book.'

book = recommend_book('fantasy').parse()
print(f'{book.title} by {book.author}')

v1 used response_model= parameter. v2 uses format= parameter and .parse() method. Response access changed from .content to .text().

Quickstart

Mirascope v2 unified llm module quickstart.

# pip install 'mirascope[openai]'
from mirascope import llm
from pydantic import BaseModel

# Simple call
@llm.call('openai/gpt-4o-mini')
def recommend_book(genre: str):
    return f'Recommend a {genre} book.'

print(recommend_book('fantasy').text())

# Structured output
class Book(BaseModel):
    title: str
    author: str

@llm.call('openai/gpt-4o-mini', format=Book)
def recommend_structured(genre: str):
    return f'Recommend a {genre} book.'

book = recommend_structured('fantasy').parse()
print(f'{book.title} by {book.author}')

view raw JSON →