Mapping the Machinery: How Two Sites Reverse-Engineered Claude Code's Source
🤖 Claude Code Unpacked by Zakaria O. I. A. · Claude Code Leaks by Abhishek Tiwari
In early 2026, Claude Code’s TypeScript source was published by accident. A cli.js.map sourcemap shipped to npm, and anyone with a browser’s devtools could read the whole thing. 1,884 files. 50+ tools. Dozens of unreleased features.
Two people turned that accident into something worth reading.
Claude Code Unpacked (ccunpacked.dev)
Zakaria built an interactive explorer that walks you through Claude Code’s internals like a museum exhibit. It covers the agent loop step by step, maps every tool and slash command by category, and documents hidden features that never shipped.
The agent loop section is the highlight. It shows what happens from keypress to rendered response: input parsing, history assembly, system prompt construction, API call, tool selection, loop iteration, rendering, hooks. Each step has source file references and a visual diagram.
The tool catalog is surprisingly useful for anyone building with AI agents. Every built-in tool is listed with what it does: file operations, execution, search, agents, planning, MCP, system, experimental. You can click through to see how Anthropic implements WebFetch, or how the Task tool spawns sub-agents.
And then there is the hidden features section. Buddy (a virtual pet that lives in your terminal). Kairos (persistent mode with memory consolidation). UltraPlan (long planning sessions on Opus-class models). Coordinator Mode (lead agent spawning parallel workers). Bridge (control Claude Code from your phone). Daemon Mode. UDS Inbox. Auto-Dream.
Some of these have shipped since the analysis. Others are still sitting behind feature flags.
Claude Code Leaks (ccleaks.com)
Abhishek Tiwari took a different approach. His site is an investigative archive with sections for Leaks, Architecture, Audit, and News. It documents 8 unreleased features, 26 undocumented commands, 32 build flags, and 120+ environment variables.
The architecture section maps the entire source tree with annotations. The audit section analyzes security implications, supply chain risks, and changes between versions. The news section tracks developments in the broader AI agent ecosystem — Codex pricing changes, Gemini Chrome Skills, the Vercel breach, alignment research.
It is more encyclopedic than exploratory. Less flash, more depth.
Why This Matters
If you use Claude Code (or any AI coding agent), these sites are worth studying. Not for the gossip — for the architecture. Seeing how Anthropic structures the agent loop, handles tool permissions, manages context windows, and orchestrates sub-agents is practical education.
The hidden features list is also a roadmap. When you see what Anthropic considered important enough to build but not release, you learn what they think the future of AI coding looks like. Coordinator Mode and Bridge tell you more about where agents are going than any blog post.
Both sites are unofficial, based on publicly available source, and may contain inaccuracies. But they are the closest thing we have to a technical tour of how a production AI coding agent actually works.
Crepi il lupo! 🐺