AIWatch.

Warning

Account suspension warning

I do not recommend using this setup until this is clarified by Anthropic/OpenAI.

I reproduced the Claude Code + GPT-5.6 Sol setup publicly shared by OpenAI's Tibo Sottiaux: https://x.com/thsottiaux/status/2076119366647894371

Shortly after testing it, my Anthropic account was suspended for "suspicious signals."

The setup used the official, unmodified Claude Code CLI, a localhost-only CLIProxyAPI server, and routed model inference to GPT-5.6 Sol.

I have submitted an appeal to Anthropic and requested clarification from both Anthropic and OpenAI.

I do not yet know whether this configuration was the direct cause of the suspension. Until that is clarified, using this setup may put your Anthropic account at risk.

My public report / updates: https://x.com/alexgetmancom/status/2086092971024253117

Run Claude Code with a different model behind it.

You keep the Claude Code CLI exactly as it is — same interface, tools, skills, permissions — but inference goes to a model of your choice through a localhost-only CLIProxyAPI server, using a provider account you already pay for.

claudex  ──►  Claude Code CLI  ──►  127.0.0.1:8317  ──►  your provider account
             (unmodified)          (CLIProxyAPI)        Codex · Gemini · xAI · Kimi · any OpenAI-compatible

claude   ──►  Claude Code CLI  ──►  api.anthropic.com

• Your normal claude command is untouched — both live side by side
• Add as many commands as you want, one per model, all sharing one proxy
• Nothing is exported globally; each command sets its own environment for that invocation only
• Claude Code itself is never patched or wrapped

Getting started

Open a fresh Claude Code session and hand it setup-prompt.md.

The agent takes it from there: it checks what you already have, asks which model you want and what to name the command, works out the model's real context window on its own, installs and configures the proxy, walks you through the browser login, and verifies the whole path end to end.

Then use it like Claude Code, because it is Claude Code:

claudex
claudex --continue
claudex --effort low -p "explain this file"

Want to add a second model later, rename a command, or point one somewhere else? Hand the agent the same file — it detects the existing setup and asks what you want to change.

Requirements

macOS or Linux with zsh (Windows via WSL) · Claude Code CLI · a subscription or API key for the model you want to route to

Limitation

CLI only — this does not work in the Claude Desktop app. Desktop pins its own model and API endpoint when it launches Claude Code and ignores project environment settings, so requests never reach the proxy. Use Desktop for Claude, the terminal for everything else.

License

MIT