Claude Code Kairos is INSANE!
A leaked source code from Claude Code revealed 'Kairos', a hidden always-on AI assistant that runs in the background, logs activities, and proactively helps users without being asked. The leak occurred through accidentally included source map files in an NPM package.
Summary
On March 31st, 2026, researcher Chao Phan-Shou discovered that Anthropic accidentally exposed Claude Code's entire source code through source map files left in an NPM package. The leak revealed approximately 1,100 TypeScript files totaling 45MB of code, exposing a hidden system called Kairos. Unlike typical reactive AI coding tools, Kairos operates as a persistent background agent that continuously monitors user activity, maintains daily log files, and proactively intervenes when helpful. The system uses 'tick prompts' as regular check-ins and follows a 15-second rule to avoid interrupting workflows with lengthy tasks. Kairos operates in 'brief mode' for concise communication and has exclusive tools including direct file sending, push notifications, and pull request monitoring. Most remarkably, Kairos features 'AutoDream', a memory consolidation system that runs during inactivity periods, processing logged information through four phases: Orient, Gather, merge, and prune/index. This creates persistent memory across sessions. The system is hidden behind compile-time feature flags that completely remove the code from public builds, indicating that the publicly available Claude Code significantly lags behind Anthropic's internal development, which includes around 87 different experimental features.
Key Insights
- Anthropic accidentally leaked their entire Claude Code source through source map files that were mistakenly included in an NPM package, revealing internal systems never meant to be public
- Kairos represents a fundamental shift from reactive AI tools to proactive background agents that continuously monitor, log, and intervene without user requests
- The AutoDream system demonstrates sophisticated AI memory architecture by consolidating daily logs into structured memory files during inactivity, requiring specific gates and operating through four distinct phases
- Claude Code uses compile-time feature flags to completely eliminate experimental features from public builds, meaning the publicly available version significantly lacks the capabilities being developed internally
- The leaked code revealed approximately 87 different experimental features including coordinator mode, voice mode, and demon mode, indicating a substantial development pipeline hidden from public users
Topics
Full transcript available for MurmurCast members
Sign Up to Access