A Proxy Lets Codex Run Any Model, Not Just OpenAI's

A new open-source proxy called opencodex sits between OpenAI’s Codex and whatever model a developer actually wants to run. Point Codex at a local address instead of OpenAI’s servers, and opencodex converts each Codex Responses API call into the format a given provider expects, then translates the response back. The project, maintained by a developer using the handle lidge-jun, ships as an npm package and installs with a single command.

The pitch is straightforward: keep the Codex CLI, App, or SDK workflow developers already know, but swap the model underneath it. opencodex ships five protocol adapters, covering Anthropic’s Messages format, Google’s Gemini format, Azure, OpenAI’s own Responses API passed straight through, and generic OpenAI-compatible chat completions. Together those adapters connect to more than 40 providers without extra configuration, per the project’s documentation, including Anthropic’s Claude, Google’s Gemini, xAI’s Grok, Zhipu’s GLM, DeepSeek, Moonshot’s Kimi, Alibaba’s Qwen, and local runtimes such as Ollama and vLLM. A developer picks a target through a provider and model syntax, running a coding task against Claude Opus in one terminal tab and against a model on their own machine in the next, with no change to how Codex itself is invoked.

The same daemon also serves Anthropic’s Messages API directly, so opencodex can route Claude Code, Anthropic’s own coding agent, through the identical roster of providers. One local process becomes the switching layer for two of the industry’s most widely used coding agents at once.

Setup is a local proxy a developer runs on their own machine, not a hosted service. The install path is npm install -g @bitkyc08/opencodex, followed by an interactive ocx init that writes a config file and wires the proxy into Codex’s own settings, then ocx start. The tool can register itself as a background service, using launchd on macOS, systemd on Linux, or Task Scheduler on Windows, so it survives reboots, and stopping it restores Codex to its original configuration with nothing left behind.

The project also builds in ChatGPT account pooling. Developers juggling several Codex accounts can have opencodex track each account’s usage window and automatically send new sessions to whichever account has the most quota left, while pinning conversations already in progress to the account that started them.

The more interesting story here is what this does to switching costs. Codex ships built to run OpenAI’s models by default. opencodex does not remove that default so much as route around it, letting a developer keep a familiar interface while swapping out the model answering behind it. That is the same pattern that turned OpenAI-compatible chat completion formats into a de facto standard: once enough tooling speaks one API shape, any provider willing to mimic that shape can plug into an entire existing ecosystem instead of building its own integrations from scratch.

That workaround carries a cost the project’s own material only gestures at. opencodex describes itself as an independent, community-maintained effort with no affiliation to OpenAI, Anthropic, or any provider it connects to, which means its translation logic depends on reverse-engineering how a first-party API behaves rather than any guaranteed contract with the vendor. The project’s disclaimer goes further: it warns that certain providers, with Anthropic named specifically, could suspend or limit accounts found routing calls through an unofficial proxy, and that using it against a provider’s own terms is a risk the operator absorbs alone.

Teams considering routing production coding work through opencodex should treat it as a bridge, not a foundation. An unofficial proxy tracking a first-party API breaks whenever that API changes shape, and any team depending on it needs a fallback plan for the day an update stops being optional.

Published on GitHub by the opencodex project on July 21, 2026.