OpenCodex is a local proxy that lets Codex, Claude Code, Claude Desktop, and Grok Build run on models from other providers. It translates each client's wire format (Codex's Responses API, Anthropic Messages) into whatever the target provider speaks, carrying streaming, tool calls, reasoning tokens, and images in both directions, with no patches to the clients themselves. Five adapters cover Anthropic Messages, Google Gemini, Azure, an OpenAI Responses passthrough, and OpenAI-compatible Chat Completions endpoints, for 40+ built-in providers.
Routed models appear in each client's own model picker, including Codex's sub-agent picker, and a request targets a model with a provider/model identifier. Combos give one virtual model ID that fails over or round-robins across providers, and an optional JEV Auto combo lets an external or self-hosted decision model pick the model and reasoning effort for each call. For non-OpenAI models, sidecars add web search and image understanding over a ChatGPT login.
OpenCodex can also pool several ChatGPT accounts for Codex authentication, refreshing their quotas and sending new sessions to the lowest-usage healthy account while existing threads stay pinned to the account that started them. A web dashboard manages providers, OAuth logins, model selection, and accounts and shows a live request log. It runs as a CLI with a background service on macOS, Linux, and Windows, as a desktop app with a tray, or as a self-hosted Docker hub, and stopping it restores Codex's original configuration.
Features
- Multi-client proxy: serves Codex, Claude Code, Claude Desktop, and Grok Build from one proxy
- Protocol translation: streaming, tool calls, reasoning tokens, and images translated in both directions
- Provider coverage: 40+ built-in providers including OpenAI, Anthropic, Gemini, xAI, Kimi, Azure OpenAI, Ollama, DeepSeek, Mistral, and any OpenAI-compatible endpoint
- Native model pickers: routed models show up alongside native ones, with reasoning effort levels
- Combos: one virtual model ID with failover or weighted round-robin across providers
- Sub-agents on any model: routed models available in Codex's sub-agent picker with fallback chains
- OAuth login: signs in with xAI, Anthropic, or Kimi accounts, forwards a ChatGPT login, or uses an API key
- ChatGPT account pool: quota-aware session routing across accounts with thread affinity
- Search and vision sidecars: web search and image understanding for non-OpenAI models
- Dashboard: providers, OAuth status, model selection, and a live request log with cache token counts
- Bounded memory: every long-lived cache and buffer has a byte budget or cap, inspectable through an admin endpoint
- Clean exit: stopping the proxy restores Codex to its original configuration
