// the find
lidge-jun/opencodex
Universal provider proxy for OpenAI Codex & Claude Code — use any LLM (Claude, Gemini, Grok, DeepSeek, Ollama…) with Codex CLI, App, SDK, and Claude Code
OpenCodex is a local proxy that translates the OpenAI Responses API (used by Codex CLI/App/SDK) into whatever format other LLM providers speak, so you can point Codex, Claude Code, Claude Desktop, or Grok Build at Claude, Gemini, Grok, DeepSeek, Ollama, or 40+ other backends while keeping the native client UI. It also manages pools of ChatGPT accounts for Codex auth, auto-routing new sessions to whichever account has quota left. Aimed at developers who like their existing CLI coding tool but want to swap the model underneath it.
It translates the actual Responses API semantics — streaming, tool calls, reasoning tokens, images in both directions — rather than just proxying chat completions, which is where most of these shims break. Operational details are taken seriously: /healthz and /readyz with documented exit codes and retry/backoff, plus a memory-ownership model where every long-lived cache has a byte budget, eviction policy, and a /api/system/memory endpoint to inspect it live. The Docker build pins the base image by digest, runs non-root, and verifies a SHA-256 manifest of the build context against copied files to catch supply-chain tampering before it ships.
The ChatGPT account-pooling feature exists specifically to spread usage across multiple logins and dodge per-account rate limits — the README's own 'use at your own risk' disclaimer about provider ToS violations is a tell that this is the part most likely to get someone's account suspended. 40+ provider adapters is a wide surface to keep in sync with upstream API changes, and the project ships preview tags (`ocx update --tag preview`) rather than a clearly stated stable release cadence, so breaking changes between providers are plausible. Routing all your coding-agent traffic (including prompts and code) through a third-party local proxy to providers like Anthropic is exactly the pattern Anthropic's own terms discourage, which is a real adoption cost beyond the technical risk. The Windows build isn't code-signed yet, so SmartScreen throws a warning on first run — a rough edge if you're installing this for someone less comfortable clicking through security prompts.