Chatgpt Codex Proxy

ChatGPT Codex Proxy
Run Claude Code on your ChatGPT Plus/Pro subscription — zero workflow change.

- 한국어 README

What is this?

This proxy lets Claude Code talk to ChatGPT's Codex backend instead of Anthropic's API.
You keep using claude exactly as before — same UI, same slash commands, same MCP tools — while inference is served by GPT.

When to use this

Why this, not ChatMock or similar tools

ChatMock and similar projects expose an OpenAI/Ollama-compatible API — great for general-purpose clients, but you'd have to switch away from Claude Code entirely.

This proxy is built specifically for Claude Code:

- Zero workflow change — claude command, keybindings, CLAUDE.md, slash commands, all work unchanged.
- MCP tools cross the boundary — Claude Code's MCP servers (Stitch, Linear, Chrome DevTools, etc.) are normally only available to Claude's backend. This proxy reads ~/.claude.json at startup, connects to your MCP servers, and injects their tool schemas into every GPT request. GPT gets the same tool access Claude would have.
- Parallel tool call safety — the proxy detects mutating tools (Edit, Write, Delete, Bash) and automatically disables parallel_tool_calls to prevent unsafe concurrent file operations.
- Claude model name passthrough — use --model claude-sonnet-4-20250514 as usual; the proxy maps it to the right Codex model automatically. Or use GPT model names directly in passthrough mode.

Example session

Features

- Anthropic Messages API compatible (POST /v1/messages)
- OAuth 2.0 login — no API key, just your ChatGPT subscription
- Full request/response transformation (Anthropic ↔ Codex)
- SSE streaming
- Claude → Codex model mapping with env overrides
- MCP tool injection (see below)

Quick start

Then in another terminal:

That's it. Claude Code is now…