claude-code-guide: A Chinese-Language Guide to Claude Code, Including a Third-Party API Relay

What claude-code-guide actually is, and who it is written for

The repository contains one file, README.md. There is no source tree, no package manifest, no test directory, and no release has been published. The default branch is main, the repository is not archived, and the last push was on 2026-04-01. Treat it as a written guide, not as an installable tool: nothing in the repository is meant to be imported or run as a dependency.

The problem it addresses is specific. Claude Code is Anthropic's command line agent, and the guide states that Claude restricts access from mainland China, so reaching anthropic.com or the documentation at code.claude.com from that region may fail, and downloading the tool or using the API may require a proxy. The README's answer is a third-party API relay, and the rest of the document is a setup walkthrough for pointing Claude Code at it.

The intended reader is a developer in that situation who is comfortable in a terminal. The README says outright that the tool is aimed at developers and is not suitable for non-technical users, and that it requires some command line familiarity. If you are outside the restricted region and already have direct API access, this guide's main recommendation does not apply to you, though its installation and settings.json sections still describe the tool accurately.

The agent loop the guide describes, and what CLAUDE.md does inside it

The README frames Claude Code as an agentic assistant rather than a completion engine. The distinction it draws is that you give a natural language instruction, and the tool plans the steps and carries them out: it reads the codebase, writes code, runs tests, and integrates the result. The concrete capabilities it lists are reading an entire codebase, running bash commands, inspecting git history, and executing tests.

Project memory is handled through a file called CLAUDE.md, described in the README as a Markdown file where project details such as code style and common problems are recorded so the tool remembers them. That is the mechanism by which a session carries conventions forward instead of rediscovering them.

On safety, the README states that Claude Code runs in a sandbox environment and supports audit logs, and that its permission policy is conservative by default: operations that could modify the system need user confirmation. Permissions can be adjusted with the /permissions command, by editing settings.json, or per session with the --allowedTools flag. The guide also mentions a Skills module for reusable workflows with git version control and marketplace sharing, and the Claude Agent SDK for building custom agents. Those are described in a paragraph, not documented with examples, so anyone who needs their exact behavior will have to look elsewhere.

Installing Claude Code and confirming it runs

The README lists the requirements first: macOS 10.15 or later, Ubuntu 20.04+/Debian 10+, or Windows through WSL; 4GB of RAM or more; Node.js 18 or higher; and an internet connection for authentication and AI processing. The npm install is the original path, and the guide warns against running it with sudo because of permission problems and security risk.

npm install -g @anthropic-ai/claude-code@latest

The README also describes a newer client that does not depend on Node.js, installed differently per platform. On macOS and Linux the shell installer is used, on macOS Homebrew is an option, and on Windows the PowerShell installer is used.

curl -fsSL https://claude.ai/install.sh | bash

brew install --cask claude-code

irm https://claude.ai/install.ps1 | iex

After installing, the guide says to move into a project directory and check the version. Seeing a version number printed is the signal that the install worked.

cd your-awesome-project
claude --version

For people migrating from the npm install, the README says to close all Claude Code sessions first, then remove the old global package so it does not shadow the native binary. Note the tense in the guide itself: it says the command to download and replace the binary is typed in the terminal, but the only command it actually prints is the uninstall.

npm uninstall -g @anthropic-ai/claude-code

Pointing Claude Code at a relay through settings.json

This is the part the guide exists for. Configuration lives in two places: ~/.claude/settings.json for user-level settings and .claude/settings.json inside a project. The README's example sets the API key, the base URL, an output token ceiling, a traffic flag, and the model name under an env object, with an empty permissions object alongside it.

{
  "env": {
    "ANTHROPIC_API_KEY": "您的APIkey",
    "ANTHROPIC_BASE_URL": "https://api.yixia.ai/",
    "CLAUDE_CODE_MAX_OUTPUT_TOKENS": 64000,
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": 1,
    "CLAUDE_MODEL": "Claude模型名称,例如：claude-4-sonnet"
  },
  "permissions": {
    "allow": [],
    "deny": []
  }
}

The key that matters is ANTHROPIC_BASE_URL. Changing it is what redirects the client away from Anthropic's own endpoint. The README's process for obtaining the key is to register at api.yixia.ai, open the token management page from the left menu, click add token, name it, select the default group, check the unlimited quota box, submit, and copy the generated key. It also offers a shell script that installs Node.js, jq, and Python 3 via Homebrew, apt, yum, or pacman depending on the platform, then writes the settings file. To use it you edit the API_KEY variable near the top, save the file as claudecode.sh, and run it.

chmod +x claudecode.sh
./claudecode.sh

Separately, additional working directories can be granted so the agent can read outside the project root, which is useful when the code you are asking about lives in a sibling checkout.

{
  "permissions": {
    "additionalDirectories": ["../docs/", "../shared/"]
  }
}

Where this guide is thin, and when it is the wrong thing to follow

The guide's central recommendation is a third-party relay, and that is its weakest point. The repository does not operate api.yixia.ai, cannot vouch for it, and has no control over whether it stays up or what it does with the prompts that pass through it. Redirecting ANTHROPIC_BASE_URL means your API key and your code context travel to a host you selected on the strength of a README paragraph. The README does not discuss what happens to that traffic, does not describe a fallback if the relay becomes unavailable, and does not document rollback to the official endpoint. For a team working on anything proprietary, that gap is the deciding factor, not the convenience.

There are smaller problems. The document claims a five-fold speed increase and attributes it to research without naming a study, so it should not be repeated as a measured result. It describes the Skills module and the Claude Agent SDK in a sentence each, which is not enough to build against. The migration section refers to a claude install step without printing it. The README is also truncated in the repository view, so the tail of the script and any later sections are not visible.

Finally, the guide is Chinese-language by design. If you need English documentation, this is not the source, and the official docs at code.claude.com are the place to look, subject to the access problem the guide describes.

Claude Code against GitHub Copilot: the difference the guide draws

The README names GitHub Copilot as the comparison point and states the distinction plainly: Copilot is oriented toward completing code, while Claude Code works at the level of the whole project. The mechanisms behind that claim are the ones listed earlier, reading the full codebase, running bash, inspecting git history, and running tests. A completion tool answers inside the file you are editing; an agent that can run commands and read history can act on the repository as a unit.

That difference cuts both ways. The guide itself concedes the cost: the tool leans toward developers, does not suit non-technical users, and needs command line familiarity. A completion plugin imposes almost none of that. It also notes that the permission model defaults to asking before anything that could modify the system, which is the right default for an agent with shell access but adds a confirmation step to workflows that a completion tool never interrupts.

If your work is mostly writing new functions inside one file, the agentic model buys you little and costs you setup. If your work involves understanding an unfamiliar repository, changing several files together, and checking that tests still pass, the model the README describes is the one that fits.

Editorial conclusion

Read this repository if you want a Chinese-language walkthrough of Claude Code installation and of the settings.json keys that redirect its API traffic; skip it if you need English documentation, versioned releases, or a maintained codebase, since the repository holds only a README and its last push was on 2026-04-01. Before following the relay instructions, verify that the third-party endpoint you intend to use is one you trust with your prompts and your key, because the guide recommends the relay but does not operate it and the README does not document rollback if the relay stops working.

Frequently asked questions

Official sources

1. claude-code-chinese/claude-code-guide on GitHub
2. Issues
3. Project website
4. README