If you've been quietly searching for "LangChain but small and typed", Mirascope has been sitting in your search results pretending to be invisible.
The Setup
Mirascope is provider-native: it doesn't abstract OpenAI and Anthropic into a fake common type, it gives you typed decorators per provider with Pydantic response models. The whole library reads like a well-mannered Python package.
{`pip install "mirascope[anthropic]"
# Pick your providers as extras — no bloat.`}The Money Pattern
Decorate a function, declare a response model, get a typed object back. Streaming, tools, and async are first-class. No Runnable, no LCEL, no DSL pretending to be a pipe.
{`from mirascope.core import anthropic
from pydantic import BaseModel
class Verdict(BaseModel):
severity: str
payout_aud: float
@anthropic.call(
model="claude-sonnet-4-5",
response_model=Verdict,
)
def triage(claim_notes: str) -> str:
return f"Decide severity and AUD payout for: {claim_notes}"
v = triage("Cracked skylight, dented gutters, Gold Coast 4217.")
print(v.severity, v.payout_aud)`}The Catch
The community is small. You won't find 12 YouTube tutorials and a Udemy course — you'll find good docs and a Discord. For some teams that's fine, for others it's a deal-breaker. Pick your tradeoff honestly.
The Verdict
Mirascope is what you'd build if you sat down to design a Python LLM library in 2026 from scratch. Less framework, more function. If PydanticAI ever gets too opinionated, this is where you're going next.