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.

Claude Code  ──POST /v1/messages──>  chatgpt-codex-proxy  ──POST /codex/responses──>  ChatGPT
             <──Anthropic response──                       <──Codex SSE response──

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

git clone <repo-url>
cd chatgpt-codex-proxy
npm install && npm run build

# Login with your ChatGPT account (browser opens)
npm run login

# Start the proxy
npm run dev

Then in another terminal:

export ANTHROPIC_BASE_URL=http://127.0.0.1:19080
export ANTHROPIC_API_KEY=dummy   # value is unused; variable must be set
claude

That's it. Claude Code is now running on GPT.

Optional: shell helper

Add this to .zshrc/.bashrc for a quick gpt alias:

gpt() {
  emulate -L zsh
  local proxy_port="${CHATGPT_CODEX_PROXY_PORT:-19080}"
  local token="${ANTHROPIC_AUTH_TOKEN:-${ANTHROPIC_API_KEY:-dummy}}"

  export ANTHROPIC_BASE_URL="http://127.0.0.1:${proxy_port}"
  export ANTHROPIC_AUTH_TOKEN="$token"
  export ANTHROPIC_API_KEY="${ANTHROPIC_API_KEY:-$token}"
  export API_TIMEOUT_MS="${API_TIMEOUT_MS:-90000}"
  export PASSTHROUGH_MODE="${PASSTHROUGH_MODE:-true}"
  unset CLAUDE_CONFIG_DIR

  echo "Using local Codex proxy on :${proxy_port}"
  claude "$@"
}

MCP tool injection

Claude Code's MCP tools are normally invisible to non-Claude backends. This proxy bridges that gap.

Claude Code                   chatgpt-codex-proxy               ChatGPT Codex API
   │  tools: [Edit, Bash, ...]      │  tools: [Edit, Bash, ...        │
   │  (deferred: stitch, qmd, ...)  │          + mcp__stitch__*       │
   │                                │          + mcp__qmd__*  ]       │
   │ ──────────────────────────────>│ ──────────────────────────────> │

How it works:

1. On startup: reads ~/.claude.json → mcpServers
2. Runs MCP handshake (initialize → tools/list) for each enabled server
3. Caches schemas in memory for the lifetime of the proxy process
4. On every /v1/messages request: appends cached tools to the Codex tools array
5. When GPT calls a tool: Claude Code receives the tool_use response, executes the real MCP call, and sends the result back through the proxy

Both HTTP (type: http) and stdio (command) MCP servers are supported.

Configuration — set PROXY_MCP_SERVERS in .env:

# Specific servers (names must match keys in ~/.claude.json)
PROXY_MCP_SERVERS=stitch,linear

# All servers registered in ~/.claude.json
PROXY_MCP_SERVERS=all

# Disabled (default)
PROXY_MCP_SERVERS=

Startup log:

[mcp-registry] connecting to: stitch, linear
[mcp-registry] stitch: 8 tools loaded
[mcp-registry] linear: 6 tools loaded
[mcp-registry] ready: 14 total MCP tools

Configuration

.env setup

cp .env.example .env

Model mapping

By default (PASSTHROUGH_MODE=true) the proxy forwards whatever model name Claude Code sends straight to Codex. Set PASSTHROUGH_MODE=false to enable automatic Claude → Codex mapping:

Override per family:

ANTHROPIC_DEFAULT_HAIKU_MODEL=gpt-5.3-codex-spark
ANTHROPIC_DEFAULT_SONNET_MODEL=gpt-5.2-codex
ANTHROPIC_DEFAULT_OPUS_MODEL=gpt-5.2-codex

Available Codex models

Shorthand aliases: gpt-5.3 → gpt-5.3-codex, gpt-5.2 → gpt-5.2-codex, gpt-5.1 → gpt-5.1-codex

Effort control

Claude Code's effort slider only affects native Claude models — it is not included in API requests when using GPT models. Control reasoning effort via:

Method 1 — model name suffix (recommended)

export ANTHROPIC_DEFAULT_SONNET_MODEL="gpt-5.3-codex-xhigh"  # xhigh
export ANTHROPIC_DEFAULT_HAIKU_MODEL="gpt-5.3-codex-spark"   # low

Method 2 — global override

PROXY_DEFAULT_EFFORT=high

Priority: thinking.budget_tokens in request → model name suffix/table → PROXY_DEFAULT_EFFORT → medium

CLI commands

API compatibility

Environment variables

Troubleshooting

# Health check
curl -fsS http://127.0.0.1:19080/health

# Port in use
lsof -tiTCP:19080 -sTCP:LISTEN -nP

# Passthrough test
gpt --model gpt-5.2

# Mapping mode test
PASSTHROUGH_MODE=false gpt --model claude-sonnet-4-20250514

# Recent logs
tail -n 120 /tmp/chatgpt-codex-proxy.log

# Tool calling smoke test
python3 scripts/tool_calling_smoke.py --base-url http://127.0.0.1:19080 --model gpt-5.2

Project structure

chatgpt-codex-proxy/
├── src/
│   ├── index.ts           # Entry point
│   ├── server.ts          # Express server
│   ├── cli.ts             # CLI commands
│   ├── auth.ts            # OAuth login
│   ├── routes/
│   │   └── messages.ts    # /v1/messages endpoint
│   ├── transformers/
│   │   ├── request.ts     # Anthropic → Codex
│   │   └── response.ts    # Codex → Anthropic
│   ├── codex/
│   │   ├── client.ts      # Codex API client
│   │   └── models.ts      # Model mapping
│   ├── mcp/
│   │   ├── config.ts      # Read ~/.claude.json MCP server configs
│   │   ├── client.ts      # HTTP + stdio MCP clients
│   │   └── registry.ts    # Tool schema cache (singleton)
│   ├── types/
│   │   └── anthropic.ts   # Types
│   └── utils/
│       └── errors.ts      # Error handling
├── .env.example
├── package.json
├── tsconfig.json
└── README.md

Security

Designed for personal, local-machine usage only.

npm run dev  # binds to localhost only
export ANTHROPIC_BASE_URL=http://127.0.0.1:19080

If you deploy to a server: add authentication, restrict CORS, add rate limiting, set token file permissions to 600, add monitoring.

Do not bind to 0.0.0.0 unless you have implemented all of the above.

License

MIT