Agent Loop docs

agentloop.providers.openai

OpenAI integration using the shared OpenAI-compatible HTTP client. Install based-models-agentloop[openai-compat] and set OPENAI_API_KEY. For automatic provider selection and retries, see the providers guide.

Constants

• BASE_URL = "https://api.openai.com/v1"
• DEFAULT_MODEL = "gpt-5.1"
• TIMEOUT_SECONDS: connect 10, read 300, write 30, and pool 10 seconds.
• QUIRKS: OpenAI's default compatibility settings. Chat Completions uses max_completion_tokens; Responses uses max_output_tokens.
• MODEL_QUIRKS and quirks_for_model(model): select model-specific settings. For o1, o3, and o4 model families, Chat Completions omits parallel_tool_calls.

build_openai_client

build_openai_client(
    *,
    api_key,
    model="gpt-5.1",
    organization=None,
    project=None,
    api="responses",
    quirks=None,
    http_client=None,
    timeout_seconds=None,
) -> OpenAICompatClient

The builder returns a direct LLMClient named openai. organization and project add the OpenAI-Organization and OpenAI-Project request headers. quirks overrides the settings selected for the model; most applications can leave it unset. timeout_seconds overrides the 300-second read timeout while retaining the other timeout defaults.

api="responses" posts to /responses and supports hosted web search. It sends the full transcript with store: false, so conversation history stays in the local Transcript. api="completions" posts to /chat/completions and supports local tools, but no hosted tools. Passing a hosted tool directly to a Completions client raises BadRequestError; Agent rejects that combination during configuration.

Both formats return a normalized ModelResponse. The client exposes name, model, api, hosted_tools, complete(request), close(), and context manager methods. hosted_tools contains web_search for Responses and is empty for Completions.

This example sends the same request through each API and reads the response text:

import os

from agentloop import Transcript, join_text
from agentloop.provider import ModelRequest
from agentloop.providers.openai import build_openai_client

transcript = Transcript()
transcript.add_user_message("Reply with exactly: ready")
request = ModelRequest(messages=tuple(transcript.messages), max_tokens=500)

for api in ("responses", "completions"):
    with build_openai_client(
        api_key=os.environ["OPENAI_API_KEY"], model="gpt-4.1-mini", api=api
    ) as client:
        response = client.complete(request)
        print(client.api, join_text(response.content))

Output:

responses ready
completions ready

Pass an httpx.Client as http_client to configure a trust store, proxy, or transport. The client remains yours to close, and close() leaves it open. Passing another HTTP client type raises ConfigurationError. The builder does not add retries; use build_client() or RetryingClient if you need them.