The Anthropic company inadvertently released the complete unobfuscated source code for the Claude Code toolkit in TypeScript as part of the NPM package claude-code 2.1.88. The source code, which was not stripped of internal components, was included in the release within the ‘map’ file cli.js.map used for debugging purposes.
To prevent similar leaks, developers are advised to add the masking “*.map” to the “.npmignore” file. The leaked code is 59.8 MB large, contains 1884 files, and comprises more than 500 thousand lines of code. Despite efforts by enthusiasts to replicate the code on GitHub, Anthropic initiated sending DMCA notices to block repositories hosting the leaked code under the US Digital Copyright Act.
Upon analysis of the leaked information, uncovered 8 upcoming features, 26 hidden commands, 32 assembly flags, 22 private repositories, and over 120 confidential environment variables within the Claude Code toolkit. Some of the hidden features discovered include:
- An “Undercover” mode that conceals AI involvement when editing public repositories.
- An “Easter egg” feature providing users with a unique virtual pet displayed alongside the command line.
- Flag CLAUDE_CODE_COORDINATOR_MODE=1 for task distribution among AI agents.
- Intersession IPC interface for messaging between AI agents.
- Proactive mode with a persistent KAIROS assistant.
- Background mode for processing prompts in the background.
- ULTRAPLAN mode for complex problem-solving.
- Auto-Dream mode for idea generation between sessions.
- Internal model name “capy