Connect Codex

Route requests through the ChatGPT/Codex subscription you already pay for.

Read as Markdown

Codex Hosted turns the ChatGPT/Codex subscription you already pay for ($20-200/mo) into an OpenAI-compatible endpoint. Work that would bill per token on the API runs on your subscription, and ProxyLLM tracks what it would have cost so you can see the savings.

Connect your subscription

Two ways, same result:

  • Dashboard: open Seats and follow the Connect Codex steps.
  • CLI: run this on the machine where you’re signed into Codex:
npx proxyllm codex connect

The CLI signs you into your ProxyLLM account in the browser, uploads your Codex sign-in, and confirms the connection. Check it any time with npx proxyllm codex status.

Models

Codex serves its own model namespace, separate from the OpenAI API names, and OpenAI rotates it over time. Treat model IDs as changeable:

  • Use the stable tiers flagship (best available) or mini (cheaper tier) as the model in your requests. ProxyLLM resolves them to your account’s current models at request time.
  • GET /v1/models lists the model names and tiers the gateway accepts. If you send a Codex model name that OpenAI has since retired, ProxyLLM substitutes the current flagship and the request succeeds.

Usage limits

Your ChatGPT plan enforces its own limits in 5-hour and weekly windows. ProxyLLM surfaces them:

npx proxyllm usage

or on the Usage page. When a window is exhausted, a routing key with more lanes falls through to the next one automatically (how routing works).

If it stops working

A Codex sign-in can expire. If requests start failing with a “no session” error, reconnect with npx proxyllm codex connect or from the Seats page.