A local proxy convinces the ChatGPT desktop app it is still talking to OpenAI

Ask the desktop app for its model list and the proxy fans out. It queries every provider in its config, merges the replies, and returns a single list. In the XDA writer's setup the locally served GLM-5.3-Flash turned up in the Codex picker as `vllm/glm-5.3-flash`, sitting next to models from his OpenCode subscription ^[8]. Routing happens after selection: once a message is sent, the proxy decides where it goes ^[9].

Redirecting Codex is cheap because Codex speaks the OpenAI Responses API and takes its base URL from an environment variable ^[5]. Translating between that wire format and whatever the target model expects is harder. Codex also has its own tool-calling surface: freeform `apply_patch`, shell access, MCP namespaces, and `tool_search` ^[25]. A model trained against a different convention emits the wrong shape. According to the dev.to writeup relaying the XDA test, MiniMax M3 got lost, while GLM-5.3-Flash and Qwen 3.8 27B worked normally ^[26]. Two of the three models named as serving the app worked ^[29]. For the ratio to mean anything in your setup you would need his runtimes, his prompts, and the Codex build he was on.

The diagnostic he offers is worth copying: if a model works in Claude Code or OpenCode but fails in Codex, check whether it is calling tools correctly before looking anywhere else ^[27].

The stated reason for a proxy instead of a fork is maintenance. Format translation sits outside the client, so it survives Codex updates ^[10]. A fork has to be re-patched on every release, and one proxy covers the CLI and the desktop app at the same time ^[11]. The XDA author says he has taken several Codex updates since and it still works ^[12].

Setup is `npm install -g @bitkyc08/opencodex` on Node 18 or later, then `ocx start`, which brings up the proxy and a dashboard on port 10100 ^[13]. `ocx init` asks which provider you want and offers to write `config.toml` for you, `ocx service` runs it in the background, and the command set is documented at opencodex.me ^[15][16]. The same author ran models on several machines across a private network behind one instance ^[20]. The dashboard reports per-call token counts and an estimated cost per model ^[22].

All of the hands-on detail traces to that one tester. The dev.to post carrying it is dated 19 September 2026, credited to Nokka, and says the draft came from the deepseek-v4.1-flash model on ollama-cloud through Nous Research's Hermes Agent before Nokka checked and edited it ^[28]. It counts four pitfalls drawn from the XDA article and the official README ^[24], and the second of them, on web search, breaks off mid-sentence in the text as published ^[30].