Skip to main content
aifeed.dev the frontpage of AI
0

Claude Code Unpacked Maps the Full Agent Internals

A developer named Zackautocracy built an interactive documentation site from the leaked Claude Code source, mapping the full agent architecture across 1,000+ files - the agent loop's 11 processing stages, 50+ tools, 70+ slash commands, and a catalog of unreleased feature-flagged systems. The hidden features are the interesting part: Kairos for persistent cross-session memory, UltraPlan for extended Opus-level planning, Coordinator Mode for parallel task distribution, Bridge for remote phone and browser control, and Daemon Mode for background execution. The site effectively turns a packaging accident into a public reference architecture. Anyone building a coding agent now has a searchable blueprint of how Anthropic structured tool invocation, permission logic, and context management - the kind of operational detail that used to take months of reverse engineering.

// 0 comments

> login to comment