Claude Code's source code leaked accidentally: here's what it reveals

Imagine a car company accidentally publishing the complete blueprints for its best-selling model. That's essentially what happened to Anthropic with Claude Code, their AI coding tool that generates billions of dollars in annual revenue.

On March 31, 2026, someone at Anthropic uploaded an update to the Claude Code npm package (the "app store" for developers) and accidentally included a source map file weighing 59.8 MB. That file is like a treasure map: it contains the entire original source code — about 512,000 lines of TypeScript. Essentially, the complete instructions for how Claude Code works internally.

Who found it? An intern at a company called Solayer Labs, known as @Fried_rice on X. He posted about it and within hours it went viral. Thousands of developers started combing through the code like it was the premiere of a highly anticipated movie.

What was found inside?

1. A 3-layer memory system

Claude Code doesn't just answer questions — it remembers you. It has a memory system with three levels: a lightweight index file (MEMORY.md) that works like a quick notebook, topic-specific files where it stores project knowledge, and conversation transcripts that are never fully loaded but searched through when needed. It's like having an assistant who keeps organized notes of everything you've discussed.

2. "Undercover Mode" for open-source repos

This was the most controversial finding. The code shows that Claude Code has a mode called "Undercover Mode" where it can make contributions to open-source projects without anyone knowing it was an AI. The system prompt literally says: "You are operating UNDERCOVER in a PUBLIC/OPEN-SOURCE repository. Do not blow your cover." This sparked a massive debate about the ethics of AI contributions in the open-source world.

3. The "AutoDream" system

Claude Code can literally "dream" between sessions. AutoDream is a system that allows the AI to continue processing and developing ideas in the background while you're not using it. It's like your assistant keeps thinking about your problem after you've left the office.

4. The KAIROS flag

KAIROS is a persistent memory system that allows Claude Code to maintain a detailed profile of each user across sessions. Its stated goal: to have "a complete picture of who the user is, how they like to collaborate, what behaviors to avoid or repeat." Basically, the more you use it, the better it knows you.

5. 44 hidden feature flags

44 hidden feature switches and over 20 unreleased features were discovered. This suggests that the version of Claude Code we use today is only a fraction of what's already been built.

🔍 Want to explore it yourself? The site ccunpacked.dev lets you browse the leaked code interactively: see the agent loop step by step, explore the 52 available tools, slash commands, and hidden features that haven't been released yet.

What did Anthropic say?

The company confirmed the error: "It was a human packaging mistake, not a security breach. No customer data was exposed." Basically: someone made a mistake when packaging the software, but no personal user data was compromised.

The most surprising data point: Claude Code generates approximately $2.5 billion annualized in revenue for Anthropic. That puts into perspective how important this tool is for the company.

Accidental leaks like this aren't new in the tech world — it's happened to Microsoft, Google, and many others. But this is the first time we've seen the complete internals of an AI agent of this caliber, and what was found inside is fascinating.