Opencodex configures a local proxy that intercepts requests from Codex/OpenAI interfaces and routes them to configured local or external model providers, effectively bringing local models into the ChatGPT ecosystem.

Using the ChatGPT Desktop app, the author runs local and open-source LLMs alongside OpenAI models via opencodex, which proxies ChatGPT Desktop to include non-OpenAI models.

Opencodex acts as an OpenAI API fake on the user’s machine, implemented as a proxy rather than a fork of the ChatGPT app, ensuring compatibility with multiple local runtimes and models.

Model compatibility and tool calls vary by model, with some like MiniMax M3 struggling with Codex-style tool calls while others such as GLM-5.3 Flash and Qwen 3.8 27B work well.

Practical caveats and recommendations include starting with GLM or Qwen before trying other models, and restarting the ChatGPT app after setup to apply changes.

Four caveats are noted: Codex tool calls may differ across models, some web searches may route through OpenAI servers unless the backend is configured, image reading for text-only models may be proxied to OpenAI, and users should verify provider terms for proxy usage.

Setup is summarized in two commands (npm install -g @bitkyc08/opencodex and ocx start) with Bun runtime included, plus an optional ocx init to auto-configure providers.

The installation is straightforward, using npm, with an optional guided ocx init that configures base URLs and model providers for runtimes like Ollama, vLLM, and LM Studio.

Web searches and vision/integration behavior can route through OpenAI servers or configured backends depending on the model, with the setup enabling flexible backends such as Anthropic, xAI, Gemini, and Exa for search.

OpenCodex supports multiple runtimes on standard ports (including Ollama, vLLM, LM Studio) and aims for compatibility with any OpenAI-compatible API or Anthropic Messages API, with official materials listing over 40 providers.

In practice, opencodex offers a seamless, user-friendly way to experiment with local LLMs inside the Codex app, noting that some models may require tool-calling adjustments.

A dashboard at localhost:10100 lets users manage providers, models, accounts, tokens, API calls, and estimated costs, and it can manage multiple ChatGPT/Codex accounts for quota control.