What is claude-code-cliproxy-advisor?

This repository is a setup guide, written in Chinese, for changing which AI models power Claude Code, Anthropic's command line coding assistant, without changing how you use the tool itself. Instead of always talking to Anthropic's own models, your requests get routed through a small local proxy program to a mix of other providers, while Claude Code's interface and workflow stay exactly the same. The guide sets up three pieces working together. Claude Code remains the assistant you actually type into. A local router program called CLIProxy sits between Claude Code and the internet, intercepting requests and forwarding them to whichever model you have actually configured, based on which built in Claude model name you selected. A small plugin, written in Go and compiled into a platform specific library file, teaches that router how to speak Claude Code's own Advisor protocol, a feature where a second model can be consulted for a second opinion, an architecture review, or a risk check before the main model gives its final answer. Under this setup, selecting Sonnet routes you to a model called GLM, selecting Haiku routes you to a model called Grok, and invoking the Advisor feature routes that specific request to a third model referred to as SOL. Everything runs on the user's own machine using API keys and login credentials the user already has for those other providers, so no new subscription is created, only local configuration. Installing Go is only needed once, to compile the plugin into a file your operating system can load, either a dylib, so, or dll file depending on whether you are on macOS, Linux, or Windows. After that one time build step, Go is not needed for daily use, and the guide provides a single shell function that launches Claude Code preconfigured to talk to the local proxy instead of Anthropic directly. The guide is unusual in that it is explicitly written as a set of instructions meant to be followed automatically by an AI coding assistant carrying out the deployment, rather than a person typing each command by hand, and it includes checks at the end to confirm every model and the Advisor feature are working correctly before declaring the setup finished.