codex-cursor-proxy

A small CLI proxy that lets you use your ChatGPT Plus/Pro subscription to access Codex models in third-party clients like Cursor. It's hacky but it works.

Prerequisites

• Bun >= 1.0
• Codex CLI authenticated with your ChatGPT account (codex must have been run at least once so ~/.codex/auth.json exists)

Quick start

bunx codex-cursor-proxy

This starts a local proxy on port 3000 and opens a public tunnel via localtunnel. On startup it prints:

OpenAI Base URL for Cursor: https://<subdomain>.loca.lt
API Key: Anything you like!

Cursor setup

1. Open Cursor settings
2. Under Models > OpenAI, set:
   • Base URL to the tunnel URL printed at startup
   • API Key to any non-empty string (e.g. x)
3. Select a model name that matches what Codex supports (e.g. gpt-5.4)

How it works

Cursor ──POST /chat/completions──> Proxy ──POST Responses API──> chatgpt.com
Cursor <──Chat Completions SSE──── Proxy <──Responses API SSE──── chatgpt.com

1. Cursor sends requests in the OpenAI Responses API format
2. The proxy sanitizes the payload (extracts system message into instructions, strips unsupported params, forces stream: true and store: false)
3. Forwards to chatgpt.com/backend-api/codex/responses using your Codex access token from ~/.codex/auth.json
4. Translates the streaming Responses API events (response.output_text.delta, etc.) into Chat Completions SSE chunks (data: {"choices":[...]})
5. Streams back to Cursor with data: [DONE] sentinel

Configuration

The proxy stores a persistent tunnel subdomain in ~/.codex/cursor-proxy/config.json so the URL stays the same across restarts.

License

MIT