// the find
huangserva/claude-code-cli
这是 Claude Code 的 CLI 客户端主体（src/ 目录），即整个终端交互层的源码。具体包含： 1. CLI 入口与命令解析 — main.tsx（4684行）、entrypoints/（CLI 模式、SDK 模式、MCP 模式） 2. 终端 UI 渲染 — components/（144 个组件，用 React + Ink 渲染终端界面） 3. 工具系统 — tools/（43 个工具，Bash、文件读写编辑、Grep、Glob、Web 搜索等） 4. API 通信层 — services/api/（与 Anthropic 后端通信） 5. MCP 协议 6. 多 Agent/Team 协作 l 7. 认证与策略 8. 辅助服务 等
This is Anthropic's Claude Code CLI source code — the actual tool, not a clone — extracted from the npm package and republished to GitHub under a 'learning purposes' disclaimer. The README explicitly states the copyright belongs to Anthropic. It covers the full terminal UI layer (React + Ink), 43 tools, MCP protocol, multi-agent coordination, and a bridge system for remote execution.
- The bridge infrastructure (30+ files under bridge/) shows how to build a robust remote-execution layer for a CLI tool — JWT auth, session management, flush gating, and transport abstraction are all there to read.
- The Ink component library is unusually mature for a terminal UI — 144 components including a full design system, custom select primitives, and keybinding management that goes well beyond what most CLI tools bother with.
- The multi-agent TaskTool and coordinator pattern is a real production implementation of parallel agent orchestration, not a toy example.
- The plugin system (commands/plugin/) handles marketplace browsing, trust warnings, and options dialogs — concrete reference for anyone building extensible CLI tooling.
- This is proprietary Anthropic code published without their authorization. The author's own README says 'copyright belongs to Anthropic, PBC.' — a DMCA takedown can and likely will arrive, making this an unreliable reference.
- No build config, no package.json, no lockfile — the repo explicitly states it 'does not include complete build configuration and dependencies.' You cannot run any of this.
- The 611 stars / 1306 forks ratio is inverted, which usually signals automated or low-quality forks rather than genuine interest — the forks are almost certainly mirrors before the takedown.
- Reading internal production code without architecture docs, test suites, or commit history gives you the what but not the why — subtle invariants and design decisions are opaque.