claude-codex

Claude Code, powered by your ChatGPT subscription.

Keep the Claude Code workflow you already know: its terminal UI, slash commands,
skills, hooks, MCP servers, and tools. claude-codex connects it to the Codex
backend using your existing ChatGPT login.

• A familiar workflow. Run the regular Claude Code CLI and pass through your usual arguments.
• Reuse your login. Load credentials from Codex CLI or OpenCode without an OpenAI API key.
• Choose your models. Configure the main model and route auxiliary requests separately, including through Ollama.

Get started

You need Claude Code, uv,
and a ChatGPT login already configured in Codex CLI or OpenCode.

git clone https://github.com/azalio/claude-codex.git
cd claude-codex
./install.sh
claude-codex

The installer creates the environment and links the launcher into ~/bin.
Make sure ~/bin is on your PATH, and keep the checkout in place.

Use it like Claude Code

# Start a session
claude-codex

# Pick up where you left off
claude-codex --continue

# Ask a question directly
claude-codex -p "Explain this repository"

The default is gpt-6.1-sol with medium reasoning.
Run /model inside a session to choose an available ChatGPT subscription model.
You can also choose the model or reasoning level when you launch:

CLAUDE_CODEX_MODEL=gpt-6.1-sol CLAUDE_CODEX_REASONING=high claude-codex

Auto-mode classifier checks use codex-auto-review with low reasoning.
Other auxiliary requests use gpt-6-luna with low reasoning.
See model configuration for overrides and Ollama routing.

How it works

Claude Code → local claude-codex bridge → ChatGPT Codex backend

The launcher starts a local proxy, translates Anthropic Messages and streaming
events into the Codex Responses protocol, and runs your installed claude binary.
Routing settings apply to that session; your existing configuration files stay intact.

This is an independent bridge using ChatGPT's internal Codex backend.
Subscription limits apply, upstream changes can affect compatibility, and some
Anthropic-specific capabilities are unavailable. Auto-mode checks consume inference
quota. Check the compatibility matrix for details.

Go deeper

• Configuration and usage — authentication, models, Ollama, TLS, logs, and context recovery.
• Gateway compatibility — supported features and current limits.
• Architecture — how the bridge is built.
• Local classifier experiments — measurements and trade-offs.