codex-claudecode-proxy

A local proxy installer CLI that translates the OpenAI OAuth API into a Claude-compatible API.

One-Liner

npx -y codex-claudecode-proxy

Requirements

macOS

• Claude Code is installed
• You are logged in to Codex CLI

Linux (Ubuntu / WSL2)

• Ubuntu (including WSL2 Ubuntu) is supported.
• systemd is required — systemctl --user must work.
  • Linux environments without systemd are not supported (the installer fails fast with guidance).
• Runs in your user session as systemd user units (no sudo; installs into your home directory).
  • Optional: keep services running when you log out:

    loginctl enable-linger $USER

• Installed user units on Linux:
  • cli-proxy-api-linux.service
  • cli-proxy-api-token-sync-linux.service
  • cli-proxy-api-token-sync-linux.path
• Token sync watches ~/.codex/auth.json and mirrors the token into ~/.cli-proxy-api/auths/....

WSL2: enable systemd

1. Create or edit /etc/wsl.conf:

[boot]
systemd=true

2. From Windows PowerShell, restart WSL:

wsl --shutdown

3. Re-open your WSL distro and confirm:

systemctl --user status

Troubleshooting (Linux)

• Verify the systemd user session is available:

  systemctl --user status

• Check installed units:

  systemctl --user status cli-proxy-api-linux.service
  systemctl --user status cli-proxy-api-token-sync-linux.service
  systemctl --user status cli-proxy-api-token-sync-linux.path

• View logs:

  journalctl --user -u cli-proxy-api-linux.service -n 100 --no-pager

• If systemctl --user fails (e.g. "Failed to connect to bus"), your environment doesn't have a systemd user session.
  • On WSL2, enable systemd via /etc/wsl.conf (above) and run wsl --shutdown.
  • On non-systemd Linux environments, this installer is unsupported.

Model -> Reasoning Effort Routing

This installer configures Claude Code's Opus/Sonnet/Haiku tiers so that:

• The visible model names encode the desired reasoning effort:
  • Opus: gpt-5.4(xhigh)
  • Sonnet: gpt-5.4(high)
  • Haiku: gpt-5.4(medium)
• CLIProxyAPI rewrites all tier requests to call the same upstream model (gpt-5.4), while setting reasoning.effort to xhigh / high / medium based on the tier.

In other words: tier selection controls reasoning level, not the upstream model.

Commands

# Install (safe to re-run)
npx -y codex-claudecode-proxy

# Status
npx -y codex-claudecode-proxy status

# Start/stop manually
npx -y codex-claudecode-proxy start
npx -y codex-claudecode-proxy stop

# Uninstall: stop background services and restore Claude Code settings
npx -y codex-claudecode-proxy uninstall

# Purge: uninstall + remove installed files
npx -y codex-claudecode-proxy purge

Integrity / Safety

• Claude Code settings are configured automatically, and a backup is created before changes.
• Running uninstall removes the proxy-related Claude settings and restores the original behavior.

License

MIT