agentcast-py

Structured output for any LLM call. Validate the response, retry with the validation error as feedback, raise CastError if all retries fail. BYO LLM, BYO validator. Zero runtime dependencies.

Python port of @mukundakatta/agentcast.

Install

pip install agentcast-py
# pydantic adapter is optional:
pip install "agentcast-py[pydantic]"

Usage

import asyncio
from anthropic import AsyncAnthropic
from pydantic import BaseModel
from agentcast import cast, adapters, CastError

client = AsyncAnthropic()

async def llm(messages):
    resp = await client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=512,
        messages=messages,
    )
    return resp.content[0].text

class Diagnosis(BaseModel):
    code: str
    severity: str
    notes: str

async def main():
    out = await cast(
        llm=llm,
        validate=adapters.pydantic(Diagnosis),
        prompt="Triage: server is on fire. Output a Diagnosis JSON object.",
        max_retries=3,
    )
    print(out)  # validated dict

asyncio.run(main())

If the LLM returns prose, fenced JSON, or near-misses, cast() extracts the JSON, validates it, and retries with the validation error in the chat history -- so the next attempt has the spec. After max_retries failures, raises CastError carrying every attempt for postmortem.

Sync LLMs work too

def sync_llm(messages):
    return openai_client.chat.completions.create(...).choices[0].message.content

out = await cast(llm=sync_llm, validate=schema, prompt="...", max_retries=2)

cast() is async (so you can await your async LLM), but the llm callable can be sync or async.

API

cast(*, llm, validate, prompt, system=None, max_retries=2, on_attempt=None) -> Any

Async. Returns the validated value (whatever the validator's value field is). Raises CastError on exhausted retries.

extract_json(text) -> Any | None

Sync. Pulls the largest valid JSON value out of arbitrary text -- whole text, fenced block, or balanced substring. Returns None if nothing parseable.

adapters.shape(spec)

Tiny built-in shape checker. Same as the sibling agentvet. {"field": "str|int|...|str?"}.

adapters.fn(predicate, error_builder?)

Predicate adapter.

adapters.pydantic(model_cls)

Wraps a pydantic v2 BaseModel. Requires pip install agentcast-py[pydantic].

CastError

Carries attempts: list[dict], last_error, last_text, last_parsed for postmortem.

API differences from the JS sibling

• cast() is async def (Python's natural pattern for LLM calls).
• llm may be sync or async; both are auto-detected.
• adapters.pydantic replaces adapters.zod as the natural Python adapter.
• Keyword args throughout (max_retries=, on_attempt=).
• Validators return Python dicts: {"valid": bool, "value": ... | "error": str}.

See the JS sibling's README for the full design notes.