Claude Code Source — Architecture Study & Learning Material

Production-grade AI Agent internals, dissected for learning.

中文版

This repository contains the extracted source code of Claude Code (Anthropic's AI programming CLI) along with 18 original deep-analysis articles (bilingual: English & Chinese) dissecting its architecture. The source was discovered from publicly available repositories and is provided strictly for educational and research purposes.

All intellectual property rights of the source code belong exclusively to Anthropic, PBC.

Kept in sync with upstream. This repository tracks the official Claude Code release stream and is kept current with the latest version. Each upstream release has a corresponding update applying the user-facing, tractable subset of its changelog — see CHANGELOG.md for the per-version record of what was applied here and what remains upstream-only.

Disclaimer · Why Study This · Architecture · Deep Analysis · Run from Source · Tech Stack · Structure · License

Disclaimer

IMPORTANT: READ BEFORE USE

This repository is strictly for educational and research purposes only.

• Source: The source code was discovered from publicly available sources, including instructkr/claw-code. The run-from-source setup is based on JiaranI/start-claude-code. It contains extracted source code of Claude Code, a product by Anthropic.
• Ownership: All intellectual property rights belong exclusively to Anthropic, PBC. This repository claims no ownership, authorship, or rights over the original code.
• Purpose: Provided solely as learning material for understanding AI agent system architecture and engineering patterns — for academic study, technical research, and educational discussion.
• Prohibited uses:
  • Commercial purposes of any kind
  • Building competing products or services
  • Redistribution or repackaging of the source code
  • Any purpose that violates Anthropic's Terms of Service or applicable laws
• Analysis articles: The deep analysis articles (claude-code-deep-analysis/) are original commentary and analysis by the repository maintainers. Code snippets are included for commentary, criticism, and education purposes. These articles do not constitute legal advice.
• DMCA / Takedown: If you represent Anthropic or believe this repository infringes on intellectual property rights, please contact us — we will promptly remove the infringing content. You may also file a DMCA takedown notice through GitHub's DMCA process.
• No warranty: Provided "as is" without warranty of any kind.

Why study this codebase?

Claude Code is one of the most sophisticated production AI agent systems publicly observable. Unlike toy frameworks and demo agents, it is engineered for real-world, hours-long coding sessions with enterprise-grade reliability. Studying it reveals hard-won engineering decisions that no tutorial covers:

Architecture at a glance

                        CLI / VS Code / IDE Extension
                                    │
                                    ▼
                        ┌───────────────────────┐
                        │   EntryPoint (cli.tsx)  │
                        │   Fast-path routing     │
                        └───────────┬───────────┘
                                    │
                        ┌───────────▼───────────┐
                        │     QueryEngine         │
                        │   Session management    │
                        │   API client wrapper    │
                        └───────────┬───────────┘
                                    │
                ┌───────────────────▼───────────────────┐
                │          query() — Core Agent Loop     │
                │   ┌─────────────────────────────────┐  │
                │   │  while (true) {                 │  │
                │   │    messages = buildPrompt()      │  │
                │   │    response = stream(messages)   │  │
                │   │    tools = extractToolCalls()    │  │
                │   │    if (!tools) break             │  │
                │   │    results = executePar(tools)   │  │
                │   │    messages.push(results)        │  │
                │   │    maybeCompress(messages)       │  │
                │   │  }                              │  │
                │   └─────────────────────────────────┘  │
                └──────┬──────────┬──────────┬──────────┘
                       │          │          │
              ┌────────▼──┐ ┌────▼────┐ ┌───▼────────┐
              │ 45+ Tools  │ │ Permis- │ │ Context    │
              │ Bash, Edit │ │ sion    │ │ Compression│
              │ Glob, Grep │ │ System  │ │ 4 layers   │
              │ MCP, LSP…  │ │ 5 modes │ │ + 413 heal │
              └────────────┘ └─────────┘ └────────────┘

By the numbers

Deep Analysis Series

We produced 18 original articles with source-level deep dives into every major subsystem. Each article is available in both English and Chinese.

Enter the analysis series (EN) → ｜ 进入分析系列 (中文) →

Part 1: Core Agent Engine

Part 2: Peripheral Subsystems

Run from Source

This repository supports running Claude Code locally from source (with reduced functionality). Verified working with version 2.1.89.

Prerequisites

Quick start

# 1. Install dependencies and generate shims
node scripts/setup.mjs

# 2. Login with your Claude subscription (opens browser)
./start.sh login

# 3. Launch
./start.sh

Running from source with Claude Pro/Max subscription — Opus 4.6 (1M context) ✓

export ANTHROPIC_API_KEY="sk-ant-xxx"
./start.sh

If you have a ChatGPT Plus/Pro/Team subscription, you can use GPT-5.4 as the agent model — no API key needed:

# 1. Login with your ChatGPT account (opens browser)
./start.sh chatgpt-login

# 2. Launch with GPT-5.4
./start.sh --model gpt5.4

Running GPT-5.4 via ChatGPT subscription as the agent model

Available GPT model aliases: gpt5.4, gpt5.4-mini, gpt5.4-nano

You can also use the full format: --model chatgpt:gpt-5.4

export OPENAI_API_KEY="sk-xxx"
./start.sh --model openai:gpt-5.4

Other OpenAI-compatible providers are also supported:

export DEEPSEEK_API_KEY="sk-xxx"
./start.sh --model deepseek:deepseek-chat

export MISTRAL_API_KEY="xxx"
./start.sh --model mistral:mistral-large-latest

rem Use API key on Windows (OAuth login requires start.sh on macOS/Linux)
set ANTHROPIC_API_KEY=sk-ant-xxx
bun src/entrypoints/cli.tsx

• ripgrep: The tar extraction in the setup script may fail on Windows. Install ripgrep manually and ensure rg is on your PATH.
• Bun path: After installing Bun, restart your terminal or add %USERPROFILE%\.bun\bin to PATH.
• OAuth login: start.sh login is macOS/Linux only. On Windows, run bun src/entrypoints/cli.tsx auth login --claudeai.

Shim files created

These files are not part of the original source — they were created to make the source runnable:

How it works

Unavailable features (missing private packages)

Since feature() returns false at runtime, all feature-gated code paths are inactive:

ABLATION_BASELINE AGENT_MEMORY_SNAPSHOT AGENT_TRIGGERS BRIDGE_MODE BUDDY BUILDING_CLAUDE_APPS CCR_AUTO_CONNECT COORDINATOR_MODE DAEMON DIRECT_CONNECT DUMP_SYSTEM_PROMPT FORK_SUBAGENT HISTORY_PICKER KAIROS MCP_SKILLS MONITOR_TOOL NATIVE_CLIPBOARD_IMAGE PERFETTO_TRACING QUICK_SEARCH SSH_REMOTE STREAMLINED_OUTPUT TEAMMEM TEMPLATES TERMINAL_PANEL TORCH ULTRAPLAN ULTRATHINK VOICE_MODE WEB_BROWSER_TOOL WORKFLOW_SCRIPTS and 59 more.

Inferred build pipeline

TypeScript source
  │
  ├─ Bun bundler (bun build)
  │   ├─ Inject MACRO.* constants (--define)
  │   ├─ Resolve feature() calls → set 89 feature flags true/false
  │   ├─ Dead Code Elimination → strip disabled feature branches
  │   └─ Bundle into single JS file
  │
  ├─ Optional: Bun compile → single-file executable binary
  │
  └─ Publish to npm (@anthropic-ai/claude-code)

Troubleshooting

Want the official release?

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

Technology stack

Project structure

claude-code-source/
├── README.md                         # This document (English)
├── README.zh.md                      # 中文版文档
├── assets/                           # Images
│   ├── anthropic-claude-code.webp
│   └── screenshot.png
├── package.json                      # Dependencies (reverse-engineered)
├── tsconfig.json                     # TypeScript configuration
├── bunfig.toml                       # Bun preload configuration
├── preload.ts                        # Runtime shim (feature flags, macros)
├── start.sh                          # Launcher script
├── scripts/setup.mjs                 # One-command setup
│
├── claude-code-deep-analysis/        # 18 original analysis articles (EN + 中文)
│   ├── README.en.md                  #   Series index (English)
│   ├── README.md                     #   Series index (中文)
│   ├── 00-core-conclusion.en.md      #   Each article has .en.md + .md
│   ├── 00-core-conclusion.md
│   ├── ...
│   └── 17-hook-system.en.md
│
└── src/                              # Claude Code source (Anthropic)
    ├── query.ts                      # Core agent loop (1,729 lines)
    ├── QueryEngine.ts                # Session management (46.6 KB)
    ├── Tool.ts                       # Tool interface & registry (29.5 KB)
    ├── main.tsx                      # Main React component (4,683 lines)
    │
    ├── entrypoints/                  # Entry points (CLI, MCP, SDK types)
    ├── commands/                     # 87+ slash commands
    ├── tools/                        # 45+ tool implementations
    │   ├── BashTool/                 #   Shell execution (430 KB of safety code)
    │   ├── AgentTool/                #   Sub-agent execution
    │   ├── FileEditTool/             #   File editing with diff matching
    │   ├── MCPTool/                  #   Model Context Protocol
    │   └── ...                       #   30+ more tools
    │
    ├── services/                     # 38 service modules
    │   ├── api/                      #   Anthropic API integration
    │   ├── mcp/                      #   MCP protocol (25 subdirs)
    │   ├── compact/                  #   Context compression
    │   ├── SessionMemory/            #   Cross-session memory
    │   └── ...
    │
    ├── components/                   # 146+ React UI components
    ├── hooks/                        # 85+ React hooks
    ├── utils/                        # 564+ utility functions
    ├── permissions/                  # Permission system (5 modes)
    ├── bridge/                       # CLI ↔ VS Code integration
    ├── constants/                    # System prompt assembly & config
    ├── memdir/                       # Cross-session memory system
    ├── skills/                       # Skill/plugin system
    ├── voice/                        # Voice input handling
    └── ...

License

• Source code: All rights belong to Anthropic, PBC. No license is granted.
• Analysis articles (claude-code-deep-analysis/): Original commentary by the repository maintainers. Code snippets are for educational purposes only.

Contact: For any questions, please reach out via GitHub Issues.