Claude Code Unpacked : A visual guide
TL;DR Highlight
An unofficial visual guide analyzing the leaked Claude Code source code, covering the agent loop, 50+ tools, and undisclosed features. A great reference for developers who want to understand how Claude Code works internally.
Who Should Read
Developers who use Claude Code in production or are building their own coding agents — especially those looking to reference agent architecture design or tool system structure.
Core Mechanics
- This site analyzes the leaked Claude Code source code (~500,000 lines) and visualizes the internal process triggered by a message input in the following sequence: Input → Message → History → System → API → Tokens → Tools → Loop → Render → Hooks → Await.
- Claude Code's tool system consists of 50+ tools, categorized as: file manipulation (FileRead, FileEdit, FileWrite, etc. — 6 tools), code execution (Bash, PowerShell, REPL — 3 tools), search/web (WebSearch, WebFetch, etc. — 4 tools), agent/task (Agent, TaskCreate, TaskList, etc. — 11 tools), MCP (mcpList, McpResourceRead, etc. — 4 tools), system (TodoWrite, AskUserQuestion, etc. — 11 tools), and experimental tools (Sleep, StructuredOutput, etc. — 8 tools).
- There are also 70+ slash commands, divided into: Setup & Config (/init, /login, /config, etc. — 12), Daily Workflow (/compact, /memory, /plan, etc. — 24), Code Review & Git (/review, /commit, /diff, etc. — 13), Debugging (/status, /cost, /heapdump, etc. — 23), and Advanced & Experimental (/advisor, /voice, /desktop, etc. — 23).
- The source code contains undisclosed features not yet publicly released. 'Kairos' is a persistent mode that integrates memory across sessions and operates autonomously in the background. 'Coordinator Mode' is a multi-agent orchestration feature where the main agent decomposes tasks and spawns parallel workers in isolated git worktrees.
- Other undisclosed features include: 'Bridge' for remotely controlling Claude Code from a phone or browser; 'Daemon Mode' — a --bg option that runs sessions in the background using tmux; 'Auto-Dream,' which automatically organizes the AI's learning between sessions; and 'Buddy,' a terminal virtual pet whose species and rarity are determined by account ID.
- The source code is organized into directories such as utils (564 files), components (389 files), commands (189 files), tools (184 files), services (130 files), hooks (104 files), and ink (96 files). An architecture explorer is also provided, allowing users to click through the full directory structure.
Evidence
- "Many questioned how a 'simple TUI' could balloon to 500,000 lines of code. One commenter analyzed that '90% of this code is likely defensive programming to prevent the agent from drifting or quietly breaking things' — meaning most of the code is dedicated to frustration regexes, context sanitizers, tool retry loops, and state rollbacks to make LLM behavior deterministic. Some suggested it was 'vibe-coded without regard for technical debt,' arguing that 500,000 lines is excessive for roughly one year of development, and noting that LLMs tend to bloat generated code unnecessarily. An actual user shared their experience building a multi-agent system with Claude Code and running into token cost issues — burning through 75% of their Pro plan weekly budget far faster than expected, leading them to shift strategy: using Claude Code for complex new implementations and pasting files directly into the web interface with Sonnet for repetitive tasks on existing code. Some cynical comments argued 'you could figure this out without the source code' and that 'Anthropic's real value is the model itself — anyone can build a frontend loop.' In contrast, others were enthusiastic about the undisclosed features (especially cross-session referencing and the Claude Code spirit animal), showing a wide range of reactions. Another developer independently built a similar analysis site around the same time (brandonrc.github.io/journey-through-claude-code), and comments marveled that multiple visualization sites emerged within just a day or two of the leak — 'something unimaginable back in 2020.'"
How to Apply
- "When designing your own coding agent or LLM-based automation pipeline, referencing Claude Code's tool categorization system (file / execution / search / agent / MCP / system / experimental) can help you quickly identify what categories of functionality you need when designing your own tool catalog. If Claude Code usage costs are higher than expected, it helps to understand how the internal agent loop works and classify your tasks accordingly — using Claude Code for complex new implementations and a web interface + Sonnet combination for simple repetitive tasks or partial file edits can reduce costs. Since undisclosed features like Coordinator Mode, Kairos, and Auto-Dream are likely to be released in future versions, if you're designing multi-agent systems or long-running agents now, referencing their architecture (parallel worktree branching, cross-session memory integration, etc.) in advance can help you establish a solid architectural direction. The full leaked Claude Code source code is available at codeberg.org/wklm/claude-code, and this visualization site can serve as a map for quickly locating which files do what — useful when you want to find the source file for a specific tool or command implementation."
Terminology
에이전트 루프 (Agent Loop)A structure in which an LLM receives a user request, calls tools, checks results, and repeats the decision-making cycle. It mimics the way humans work to handle complex tasks step by step.
에이전트 하네스 (Agent Harness)A framework or wrapper code that connects an LLM model to real-world tasks. It handles model API calls, tool execution, state management, and more.
MCP (Model Context Protocol)Anthropic's protocol for enabling LLMs to communicate with external tools and data sources in a standardized way. Like a USB standard, it allows any tool to be connected using the same interface.
WorktreeA Git feature that allows the same repository to be checked out and worked on simultaneously in multiple directories. In Claude Code, it is used so that parallel agents can work independently without conflicting with each other.
드리프트 (Drift)A phenomenon where an agent deviates from its original goal and behaves in unexpected ways. It occurs due to the probabilistic nature of LLMs and requires complex state management code to prevent.
Vibe CodingA development style that uses LLMs to prioritize fast feature implementation over code quality or structure. It tends to accumulate technical debt and can lead to unnecessarily bloated code.