llm-async-codex

ChatGPT Codex subscription provider for llm-async.

from llm_async_codex import CodexProvider

provider = CodexProvider.from_codex_home()
response = await provider.acomplete(
    model="gpt-5.5",
    messages=[{"role": "user", "content": "Review this repository."}],
    stream=True,
)
async for chunk in response.stream_content():
    print(chunk, end="", flush=True)

This reads an existing Codex CLI login from $CODEX_HOME/auth.json or ~/.codex/auth.json by default (CodexProvider.from_codex_home()), or a CodexCredentials loaded from any path via load_credentials(path).

Login

If you don't already have a Codex CLI login, log in directly:

llm-async-codex login
llm-async-codex login --device-code   # headless/SSH, no local browser needed
llm-async-codex login --auth-file ~/.my-tool/auth.json
llm-async-codex login --verbose

or from Python: await login(device_code=False, auth_path=None, verbose=False).

Tool calling

Tool calling works through the normal streaming API: call provider.acomplete(..., stream=True, tools=[...]), drain response.stream_content(), then read response.main_response.tool_calls. Requires llm-async>=0.5.2.

Model catalog

await provider.list_model_slugs() lists models available to the authenticated account. await provider.get_model_capabilities(model) returns its context-window and automatic compaction limits when present.

Limitations

• Streaming only: the Codex backend requires stream=True; non-streaming requests are rejected.
• No stateless multi-turn: the backend rejects store=True ("Store must be set to false"), so previous_response_id-based continuation (as used in some of llm-async's other Responses API examples) does not work here. Resend the full conversation history (including function_call/function_call_output items) on every turn instead.
• max_output_tokens is ignored because the Codex backend rejects it.