A local router that smuggles rival LLMs into OpenAI Codex

What it does

Codex Router is a local, credential-isolating router that sits between the Codex App or CLI and a growing list of external providers. It translates requests so that Anthropic, DeepSeek, Kimi, xAI, and others speak the Responses API that Codex expects, then merges them into the native model picker as if they were first-class citizens. Your API keys are entered only through hidden local terminal prompts and never exposed in chat, while existing OpenAI logins and settings remain untouched.

The interesting bit

The project treats credential hygiene as a first-class feature: it can detect existing OAuth sessions from the Kimi and Grok CLIs, reuse them securely, and roll back changes via Git if something breaks. It also distinguishes between vendor-direct billing and reseller routes—like Ollama Cloud or Alibaba’s Qwen plan—so the same model ID can coexist under separate quotas and credentials.

Key highlights

• Merges external models into Codex’s native catalog; routed providers appear in the standard model picker alongside GPT models.
• Credential-aware catalog only surfaces models from providers with stored API keys or valid OAuth sessions.
• Reuses existing CLI OAuth sessions for Kimi and Grok instead of demanding fresh tokens.
• Supports interactive curation of unlisted models from live provider catalogs like Groq, OpenRouter, and Fireworks.
• Quota tray shows router-observed traffic for providers that lack a published balance API, such as Qwen and Google Gemini.

Caveats

• Curated models added from live catalogs are local to your machine and are not vetted by the repository’s compatibility tests.
• Providers like Qwen and Google Gemini offer no quota or rate-limit headers, so the tray can only show observed traffic rather than real account balances.
• Cross-vendor models routed through Alibaba’s DashScope use a compatibility-mode request profile because DashScope rejects native thinking parameters from the original vendors.

Verdict

Worth a look if you live in the Codex interface but want to shop around for models or keep sensitive API keys out of chat history. Skip it if you already use a generic OpenAI-compatible gateway and don’t mind managing credentials manually.