Xypro

Your ChatGPT Codex, speaking fluent OpenAI on localhost.

What it is

Xypro is a small local proxy. It lets any OpenAI-compatible app use your ChatGPT Codex subscription, with no API key bill.

You keep writing normal OpenAI code. Xypro does the translation behind the scenes.

How it works

• Your app calls the usual OpenAI endpoint at http://localhost:1455/v1.
• Xypro signs the request with your ChatGPT login and forwards it to the Codex backend.
• The reply comes back in plain OpenAI shape, streaming or not.

That is the whole trick. Your app never knows Codex is involved.

Install

npm install -g xypro

Then connect once and start it:

xypro auth     # sign in with ChatGPT, one time
xypro serve    # start the proxy, opens the control panel

Prefer not to install globally? Use npx:

npx xypro auth
npx xypro serve

That is the whole setup. Your code keeps calling OpenAI, Xypro points it at Codex.

Commands

• xypro serve starts the proxy and opens the control panel.
• xypro auth connects your ChatGPT Codex account (one time).
• xypro key [value] sets an API key your apps must send. No value means it generates one.
• xypro key clear removes the key so the proxy is open again.
• xypro help shows the help.

Use it from your app

Point any OpenAI client at the Xypro base URL. Nothing else changes.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:1455/v1", api_key="xypro")
client.chat.completions.create(
    model="gpt-5.5",
    messages=[{"role": "user", "content": "hello"}],
)

curl http://localhost:1455/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.5","messages":[{"role":"user","content":"hi"}],"stream":true}'

If you set a key with xypro key, send it as Authorization: Bearer <key>.

Models

These are the only models a ChatGPT account can use through Codex (checked against the live backend):

• gpt-5.5
• gpt-5.4
• gpt-5.4-mini

Each one takes a reasoning_effort of none, low, medium, high, or xhigh. More effort means more thinking and more tokens. none is fastest, xhigh thinks the hardest.

A couple of things Xypro handles quietly so your app does not break:

• It maps reasoning_effort: "minimal" (which Codex does not accept) to low.
• It drops max_tokens instead of forwarding it, because the Codex backend rejects output caps.

The control panel

Open http://localhost:1455 after xypro serve.

• Providers: connection status, connect or reconnect, refresh, and a quick model test.
• Quick test: fire one request to confirm a model works.
• Usage: live token consumption, totals, by model, and recent requests.
• Settings: default reasoning effort, reasoning summary, and default model.

Configuration

Settings live in proxy/.env (optional, sensible defaults otherwise):

• PORT is the port Xypro runs on. Default is 1455.
• PROXY_API_KEY is the key your apps must send. Empty means open, which is fine locally.
• CODEX_PROXY_DATA_DIR is where your token, usage, and settings are stored. Default is ~/.codex-proxy.

Your tokens never leave your machine.

Notes

• Codex always thinks first, so non-streaming replies are assembled after the stream finishes.
• Reasoning text is returned on a reasoning field next to the normal content.
• This maps text chat. Tool calls and vision are not translated yet.

Develop

Two live-reloading servers, backend on 1455 and frontend on 5173:

npm run dev

The backend lives in proxy/ (Express) and the control panel in web/ (Vite and React).

License

MIT