0xsero/codex-shim

Codex Desktop is an AI coding application that only lets you use a short list of models that its developers have approved. If you have API keys for other AI providers, such as Anthropic, DeepSeek, or others, there is no built-in way to swap them in. Codex-shim is a small local Python server that works around this limitation by pretending to be the official API that Codex connects to, then quietly routing your requests to whichever model you actually want to use. The setup works by running the shim as a background process on your own computer. Codex Desktop talks to it instead of to OpenAI's servers. The shim reads a settings file that lists your own models and API keys, translates Codex's requests into the format each provider expects, sends them off, and translates the responses back before handing them to Codex. Supported upstream providers include OpenAI, Anthropic-compatible endpoints, and a generic format used by many other providers. Your API keys stay in your own settings file and are never written into any generated catalog file. If you have a ChatGPT subscription with Codex access, the shim can also route a synthetic model slot directly to ChatGPT's servers, using your subscription quota rather than API credits. On macOS, there is an additional step required to make models actually appear in the picker inside Codex Desktop. The app has an allowlist built into its packaged files that hides any model not on a hardcoded list. The README includes step-by-step instructions for unpacking those app files, modifying one line to disable the allowlist, repacking everything, updating a security hash in the app's settings file, and re-signing the app so it will launch. The instructions also cover how to roll back the change. The shim was tested on Codex Desktop version 0.133.0-alpha.1 on a Mac with Apple silicon. Linux and Windows users can skip the app-patching section and use the shim itself without modification. The project requires Python 3.11 or newer.