Anthropic Accidentally Leaked Claude Code's Source—The Internet Is Keeping It Forever
Anthropic unintentionally leaked the full source code for its advanced Claude Code AI agent by publishing a JavaScript source map file to the npm registry, bypassing intended ignore settings. This debug file exposed 512,000 lines of internal code, including architecture for LLM API orchestration, multi-agent systems, permission logic, OAuth flows, and unreleased features. Notable components include “Kairos,” a daemon for memory consolidation, and “Buddy,” a Tamagotchi-like AI pet. The code also contained an “Undercover Mode” to avoid leaking internal project names in open-source contributions. After the leak surfaced on social media, Anthropic removed the package and began issuing DMCA takedowns. However, mirrors appeared on decentralized platforms and a developer quickly “clean-room” rewrote the architecture in Python, gaining massive attention and making removal legally difficult. The situation highlights copyright complications with AI-generated code and underscores the limits of DMCA enforcement in a decentralized context; once leaked, the code remains widely available and out of Anthropic’s full control.
