Codex Gateway

An OpenAI-compatible API powered by a ChatGPT subscription.

Overview

This gateway speaks the standard OpenAI wire format. Any library, bot, or tool that talks to OpenAI works here by changing two values — the base URL and the API key. No code changes, no SDK swap.

Base URL —

Authentication

Send your key as a Bearer token on every /v1 request:

Authorization: Bearer sk-cgw-YOUR_KEY

The Bearer prefix is optional. A bare key works too, as does x-api-key: sk-cgw-…. Keys are created in the dashboard.

Quickstart

Chat completions

POST/v1/chat/completions

The standard endpoint. Supports multi-turn history, system messages, streaming, function calling, and image inputs.

Field	Type	Notes
`model`	string	See Models. Ignored if the key is pinned.
`messages`	array	Required. `system` / `user` / `assistant` / `tool`.
`stream`	boolean	SSE chunks terminated by `data: [DONE]`.
`tools`	array	Function definitions — see Function calling.
`tool_choice`	string/object	`auto`, `none`, `required`, or a named function.
`temperature`	number	Passed through.
`max_tokens`	number	`max_completion_tokens` also accepted.

Response is the usual OpenAI shape:

{
  "id": "chatcmpl-…",
  "object": "chat.completion",
  "model": "gpt-5.4",
  "choices": [{
    "index": 0,
    "message": { "role": "assistant", "content": "Hello!" },
    "finish_reason": "stop"
  }],
  "usage": { "prompt_tokens": 12, "completion_tokens": 3, "total_tokens": 15 }
}

The model field in the response is the model that actually ran. If a key is pinned, or you sent a legacy slug that was remapped, this is how you tell.

Streaming

Set "stream": true to receive server-sent events.

Each chunk carries a delta; the stream ends with data: [DONE].

data: {"choices":[{"delta":{"content":"Hel"},"finish_reason":null}]}
data: {"choices":[{"delta":{"content":"lo"},"finish_reason":null}]}
data: {"choices":[{"delta":{},"finish_reason":"stop"}]}
data: [DONE]

Function calling

Define tools exactly as you would with OpenAI:

When the model decides to call a tool you get finish_reason: "tool_calls" and a tool_calls array. Run the function, then append the result as a tool message and call again:

{ "role": "tool", "tool_call_id": "call_abc123", "content": "{\"temp\":34}" }

Simple endpoint

POST/v1/ask

For bots and scripts that do not want the chat envelope.

{ "text": "…", "model": "gpt-5.4", "usage": { "total_tokens": 42 } }

Other endpoints

Endpoint	Purpose
POST`/v1/responses`	OpenAI Responses-style shape.
GET`/v1/models`	List available models.
GET`/health`	No auth. Liveness + link state.

Models

Model	Best for
`gpt-5.5`	Highest quality, complex reasoning
`gpt-5.4`	Balanced default; strong multilingual and Arabic
`gpt-5.4-mini`	Cheapest and fastest — classification, extraction, structured JSON
`gpt-5.3-codex`	Code-tuned
`gpt-5.3-codex-spark`	ChatGPT Pro only, research preview

Legacy slugs keep working. gpt-4o, gpt-4, gpt-3.5-turbo, o1, o3 and friends are auto-mapped to the nearest model above, so existing code runs unchanged.

Pinned keys override the request. An API key can be locked to one model in the dashboard. Requests made with it ignore the model field entirely — useful for third-party tools that hardcode a slug you cannot edit.

Errors

Errors use the OpenAI envelope, so SDK error handling works unchanged.

Status	Code	Meaning
401	`invalid_api_key`	Missing, unknown, revoked, or disabled key.
405	`method_not_allowed`	Wrong verb — the `Allow` header names the right one.
429	`rate_limited`	ChatGPT plan quota exhausted. Back off; re-linking will not help.
503	`needs_relink`	No linked ChatGPT account, or its session was revoked. Needs a human in the dashboard.
504	`timeout`	Upstream took too long.

Limits

Capacity is the ChatGPT plan's quota, shared across every key. Treat 429 as a hard backoff signal.
Requests run with bounded concurrency — bursts queue instead of failing.
The gateway is stateless: send full message history each call to keep context.
n > 1, logprobs, and embeddings are not supported.

For AI agents

Building with a coding model? Hand it /llms.txt — a plain-text spec of this whole API, written to be pasted straight into a model's context alongside your key.

Prompt to use: "Here is the API spec for my LLM gateway. Use it with key sk-cgw-… to add AI features to this project."