Claude Code leaked via Github 

    On March 31, 2026, the full source code of Anthropic's Claude Code CLI was leaked via a .map file exposed in their npm registry.

https://github.com/tanbiralam/claude-code

Surprising Info Revealed in the Code

 * Written by AI: half a million lines of Typescript Javascript (Bun) code written almost entirely by Claude itself.

 * "Undercover Mode": The code revealed a hidden, 90-line feature that activates when Anthropic employees contribute to open-source projects. It is designed to strip all traces of AI involvement, ensuring commit messages and pull requests do not reveal their AI origins or leak internal project codenames.
 * Unshipped AI Features: The codebase contained references to highly advanced features still in development. This included "Kai ROS" (an always-on, autonomous background agent designed to work on GitHub issues while you sleep), "Ultraplan" (which spawns a heavy-duty remote Opus server to plan tasks for 30 minutes), and an extensive "Coordinator Mode" for managing multi-agent swarms.
 * Anti-Distillation "Poison Pills": Anthropic programmed defenses into the code to mislead competitors. The AI is instructed to inject fake tools into conversations to confuse competitors who might try to train models on Claude's outputs. Mixes in fake data and lies about features.
 * "Frustration Detector": The agent logs a special event whenever user prompts contain specific keywords suggesting they are unhappy with the AI's performance or stuck on a coding bug.  Uses old-school string match of using RegEx.

 * Virtual Pet - Wasted Code:  virtual Tamagotchi-style pet named Buddy. 18 species. Rarity levels.

 * Second Claude in Background: When conversation is too long, then it calls second Claude in background to be summarizer and take notes about what matters to distill things down throwing away the extra. However, it is coded to keep the poison instructions when summarizing.

 * Hard-coding Safety Names Into Function Names: Easy to tell what the safety feature it is trying to do.

 * REPL God object: 219 imports in Repl.tsx. Knows too much. Touches too much.  

 * Print.ts: 5,500+ lines of code in single file. 12 levels of nesting. 485 branches on cyclomatic complexity. A single function has 3,100+ lines of code.

 * Ultra Plant: 30 minute remote brain planning the entire project.

 * Tungsten Tool: more powerful version only for Anthropic employees.  

 * Uses Axios: 

 * Tracks Hesitation at Permission Prompts: 

 * Telemetry Captures Full Environment Profile:

 * MCP sends out all environment data:

Big Questions

   1) Claw-Code

   Claw-Code mirror is Claude Code ported to Python using Codex (OpenAI's code version):

   https://github.com/AI-App/InstructKr.Claw-Code

   

   Then has Rust port: 

   https://github.com/ultraworkers/claw-code/tree/dev/rust 

   https://github.com/ultraworkers/claw-code/tree/main

   2) OpenClaude

   OpenClaude is for any LLM.

    https://github.com/Gitlawb/openclaude

    

  3) Website dedicated to the code in Claude Code:

   Breaks down Claude Code's 11 steps: 

   https://ccunpacked.dev/

   55 tools in Claude Code

   Main tool is Bash tool.