A source-first snapshot of the Claude Code src/ tree
Terminal-native AI coding, tool orchestration, MCP integration, plugin loading, permission enforcement, multi-agent collaboration, and remote control infrastructure in one codebase.
这个仓库保存的是 Claude Code 项目的 src/ 源码目录快照，重点展示其终端 AI 编码产品在实际工程中的核心实现方式。
从代码结构来看，它并不是一个单一职责的小工具，而是一套完整的智能编码运行时：既有命令行入口、React + Ink 终端 UI、多种工具调用链，也有插件系统、MCP 集成、多代理协作、权限治理、远程桥接、会话记忆、语音能力和大量工程化基础设施。
如果你想研究下面这些方向，这份源码很有参考价值：
- 如何把大模型能力做成真正可用的终端产品
- 如何在 CLI 中组织复杂的工具系统与权限系统
- 如何实现多代理/子代理协作与后台任务
- 如何在本地工程、远程会话、插件生态、MCP Server 之间建立统一抽象
- 如何用 TypeScript、Bun、React、Ink 构建大型终端交互应用
本仓库当前以 src/ 目录为主，适合作为源码研究、结构分析和二次整理使用。
- 已同步：
src/全量源码 - 已补充：
README.md - 已排除：所有 ZIP 文件（例如
src.zip不会进入版本控制）
代码中可以看到清晰的 CLI 入口与命令体系，包含大量 slash commands、终端组件、消息渲染、任务面板、状态展示和会话控制逻辑。它不是“简单调用模型 API”的脚本，而是面向真实交互场景设计的终端产品。
从 src/tools/ 和 src/tools.ts 可以看出，这套系统内置了丰富的工具能力，包括但不限于：
- Bash / PowerShell 执行
- 文件读取、编辑、写入
- Grep / Glob / Web Fetch / Web Search
- Todo / Task 系列工具
- MCP 资源读取
- LSP 工具
- AgentTool 多代理任务编排
- Plan / Worktree / Config 等运行控制工具
这意味着项目并不是把“工具”当作点缀，而是把工具调用当作核心操作模型。
AgentTool、任务模块、teammate/swarm 相关代码表明，这个系统具备把任务拆分给子代理、后台代理、远程代理或隔离 worktree 代理的能力。
这类能力对于复杂代码改造、并行执行、长任务跟踪和人机协作都非常重要。
从 src/utils/permissions/、BashTool、PowerShellTool、filesystem.ts 等模块可以看出，项目对危险文件、危险目录、只读命令校验、路径校验、规则系统、权限提示、自动模式边界等都有细致实现。
这让它更接近“可部署的安全代理系统”，而不只是实验性原型。
代码中同时存在成熟的插件加载器、Marketplace 逻辑、Hook 体系，以及 MCP Server 连接与资源管理逻辑。
也就是说，这个项目不仅能扩展“命令”，还能扩展“工具、技能、Agent、资源连接方式”，生态面非常宽。
源码中可以看到对 Remote Control / Bridge、Desktop、Mobile、Slack、GitHub App、Voice、Chrome Native Host 等能力的支持痕迹。
这说明它的设计目标并不是一个只跑在本地终端里的孤立程序，而是一个可连接多入口、多环境的智能编码平台。
flowchart TD
A[Entrypoints / CLI] --> B[Commands Layer]
B --> C[Prompt & Session Runtime]
C --> D[Tool Orchestrator]
D --> E[File / Shell / Web / MCP / LSP Tools]
D --> F[AgentTool / Task System / Background Work]
C --> G[React + Ink Terminal UI]
C --> H[Permissions / Sandbox / Policy Controls]
C --> I[Plugins / Skills / Hooks / Marketplace]
C --> J[Bridge / Remote Control / Mobile / Desktop]
C --> K[Memory / History / Telemetry / Settings Sync]
| 目录 | 作用 |
|---|---|
src/entrypoints |
CLI 与运行入口 |
src/commands |
Slash commands 与用户交互命令体系 |
src/tools |
模型可调用工具定义与执行逻辑 |
src/components |
终端 UI 组件、消息视图、对话框、任务展示 |
src/ink |
自定义终端渲染与交互基础设施 |
src/utils/permissions |
权限规则、路径校验、安全边界 |
src/utils/plugins |
插件发现、加载、验证、Marketplace |
src/services/mcp |
MCP 连接管理与运行时接入 |
src/bridge |
远程控制与桥接能力 |
src/tasks |
本地/远程/后台/多代理任务调度 |
src/services |
API、记忆、同步、策略、遥测等服务 |
src/keybindings / src/vim |
键位系统与 Vim 风格交互支持 |
src/voice / src/services/voice* |
语音输入相关能力 |
- 它展示了一个大型 AI CLI 产品如何真正落地，而不是停留在 Demo 阶段。
- 它把“命令、工具、权限、插件、MCP、代理、UI、远程控制”串成了统一系统。
- 它既有产品交互层，也有底层工程治理层，适合从多个视角学习。
- 它对做 AI Agent、Coding Agent、Terminal UX、Tool Use Runtime 的开发者尤其有参考价值。
- 想研究 AI Coding Agent 架构的开发者
- 想实现终端 TUI 产品的工程团队
- 想搭建插件化、可控、可审计 Agent 平台的项目维护者
- 想拆解 Claude Code 类产品设计思路的学习者
由于当前仓库以源码目录快照为主，README 会尽量忠实于已同步代码，不会虚构不存在的构建脚本、发布流程或额外资源文件。后续如果补充 package.json、构建配置、测试配置、示例脚本等内容，这份 README 也可以继续扩展成完整项目文档。
This repository is a source-first snapshot of the Claude Code src/ tree, focused on the implementation details that make a serious AI coding product work inside the terminal.
From the code layout, this is clearly more than a thin wrapper around an LLM API. It includes a substantial command system, a React + Ink terminal UI, rich tool orchestration, plugin loading, MCP connectivity, multi-agent execution, permission enforcement, remote control plumbing, memory/session infrastructure, voice-related modules, and a broad set of production-oriented utilities.
If you're interested in any of the following, this codebase is especially useful:
- building a real AI coding experience for the terminal
- designing a tool-driven agent runtime instead of a plain chat shell
- implementing multi-agent delegation and background task execution
- combining local workspace access, remote sessions, plugins, and MCP under one architecture
- scaling a TypeScript/Bun/React/Ink CLI application beyond prototype quality
This repo currently focuses on the src/ directory and is best treated as a code snapshot for study, exploration, and future curation.
- Included: the full
src/source tree - Added: a curated bilingual
README.md - Excluded: ZIP archives of any kind, including
src.zip
The code shows a real CLI entrypoint, an extensive slash-command surface, interactive terminal components, task/status views, message rendering, and session lifecycle management.
This is product-grade terminal software, not a minimal script that forwards prompts to a model.
The src/tools/ tree and src/tools.ts reveal a broad built-in tool ecosystem, including:
- Bash and PowerShell execution
- file read / edit / write flows
- Grep, Glob, Web Fetch, and Web Search
- Todo and task-management tools
- MCP resource access
- LSP-assisted tooling
- AgentTool for delegated execution
- plan/worktree/config/runtime control tools
That breadth makes the architecture especially compelling for anyone designing tool-using coding agents.
The presence of AgentTool, task abstractions, teammate/swarm logic, background execution, isolated worktrees, and remote agent flows suggests a system built to split, schedule, and supervise complex work rather than merely talk about delegation conceptually.
The permission-related modules indicate serious investment in path validation, dangerous file and directory protection, shell command classification, rule-based access control, approval UX, and bounded automation modes.
That makes the repository particularly valuable for developers thinking about safe agent deployment in local environments.
This codebase supports plugin discovery/loading, marketplace-facing plugin workflows, hooks, custom commands, and MCP server connectivity/resource access.
In other words, extensibility here is not limited to "add a command"; it reaches into tools, skills, agents, resources, and operational integrations.
There are clear signs of support for remote control / bridge flows, desktop and mobile touchpoints, Slack and GitHub integrations, voice-related capabilities, and browser/native-host connections.
That gives the repository the shape of a broader intelligent coding platform, not just an isolated command-line app.
flowchart TD
A[Entrypoints / CLI] --> B[Commands Layer]
B --> C[Prompt & Session Runtime]
C --> D[Tool Orchestrator]
D --> E[File / Shell / Web / MCP / LSP Tools]
D --> F[AgentTool / Task System / Background Work]
C --> G[React + Ink Terminal UI]
C --> H[Permissions / Sandbox / Policy Controls]
C --> I[Plugins / Skills / Hooks / Marketplace]
C --> J[Bridge / Remote Control / Mobile / Desktop]
C --> K[Memory / History / Telemetry / Settings Sync]
| Directory | Purpose |
|---|---|
src/entrypoints |
CLI and runtime entrypoints |
src/commands |
Slash commands and user-facing command handlers |
src/tools |
Model-invocable tool definitions and execution layers |
src/components |
Terminal UI, dialogs, task views, and message rendering |
src/ink |
Terminal rendering and interaction infrastructure |
src/utils/permissions |
Permission rules, path validation, safety controls |
src/utils/plugins |
Plugin discovery, loading, validation, marketplace logic |
src/services/mcp |
MCP connectivity and runtime integration |
src/bridge |
Remote control and bridge subsystems |
src/tasks |
Local, remote, background, and agent task scheduling |
src/services |
API, memory, sync, policy, telemetry, and operational services |
src/keybindings / src/vim |
Keyboard system and Vim-style interaction support |
src/voice / src/services/voice* |
Voice-related capabilities |
1902source files are present in the currentsrc/snapshot.86command directories are available undersrc/commands.42tool directories are available undersrc/tools.- The code strongly indicates usage of TypeScript, Bun, React, and Ink.
- The architecture spans command handling, tool execution, UI rendering, plugins, MCP, permissions, and remote workflows.
- It shows how an AI coding interface can be structured as a full software platform.
- It connects UX, orchestration, security, and extensibility in one implementation.
- It is useful both for product thinking and for systems-level engineering study.
- It is especially relevant for builders working on coding agents, local AI tooling, terminal UX, plugin platforms, and safe tool invocation.
- developers studying AI coding agent architecture
- teams building sophisticated terminal-first products
- maintainers designing auditable, extensible, tool-using agents
- engineers reverse-engineering the structure of products similar to Claude Code
This README intentionally stays faithful to the material currently tracked in the repository. Since the repo is presently centered on the src/ tree, it does not invent missing build scripts, release tooling, or packaging metadata. As the repository grows, this document can evolve into a fuller project manual.
