This repository provides a lightweight Python interface for talking to OpenAI's GPT models (including gpt-5 and gpt-5-codex). It started life as the Anthropic Claude Agent SDK; the transport layer has been replaced so the same high level API now talks to OpenAI's Responses API instead of the Claude CLI.

pip install -e .

• Python 3.10+
• An OpenAI API key (set OPENAI_API_KEY or pass it through ClaudeAgentOptions.openai_api_key)

import anyio
from claude_agent_sdk import query


async def main() -> None:
    async for message in query(
        prompt="Write a short haiku about asynchronous Python",
    ):
        print(message)


anyio.run(main)

The SDK yields fully typed message objects. In the example above the first message will be an AssistantMessage with a TextBlock containing the model's response, followed by a ResultMessage with timing and token usage details (if provided by the API).

For multi-turn interactions use ClaudeSDKClient:

from claude_agent_sdk import ClaudeAgentOptions, ClaudeSDKClient


options = ClaudeAgentOptions(model="gpt-5-codex", temperature=0.4)


async def run_conversation() -> None:
    async with ClaudeSDKClient(options=options) as client:
        await client.query("Please introduce yourself")
        async for message in client.receive_response():
            print(message)

        await client.query("Thanks! Can you explain how streaming works?")
        async for message in client.receive_response():
            print(message)


anyio.run(run_conversation)

The client keeps the conversation history in memory and re-sends it to the API for every new prompt.

ClaudeAgentOptions now exposes the most common OpenAI configuration knobs:

Legacy Claude-specific fields are still present on ClaudeAgentOptions for API compatibility but they are ignored by the OpenAI backend.

The test-suite runs without contacting the real API by injecting a mock transport. You can execute it locally with:

pytest

MIT