Claude Code's Source Code Just Leaked. Here Is What It Reveals About How You Should Be Using It.
Anthropic accidentally leaked Claude Code's full source code via npm. Here is what the architecture reveals about how most people are massively underusing it.
Two leaks in a single week. First, Business Insider reported that Anthropic accidentally exposed Claude Code's full source code through an npm package. Then a VentureBeat deep dive confirmed the scale: 512,000 lines of production TypeScript, mirrored publicly on GitHub within hours. The reaction was global - 60,000 forks before Anthropic could issue DMCA takedowns.
But here is the thing. The interesting part is not the leak itself. It is what the architecture reveals about how most people are massively underusing the tool.
It is not a chatbot
Most people treat Claude Code like a chatbot that happens to run in their terminal. Type a question, get an answer. Maybe ask it to edit a file.
The source code tells a completely different story. Claude Code is a full agent runtime with 11 architectural layers: a tool system, command system, memory system, permission engine, task manager, multi-agent coordinator and MCP client/server, all wired together under one execution pipeline.
That means the chat interface most people interact with is roughly 10% of what the tool can do. The other 90% is in the systems built around the model.
We broke down all nine architectural insights in our interactive power user guide.
Three things that change immediately
First: your CLAUDE.md file. The source confirms this is not optional documentation. It is an operating manual that gets injected into every single conversation turn. Not at session start. Every turn. You have 40,000 characters to shape how Claude Code behaves, and the memory system (MEMORY.md) means it learns from your corrections across sessions. Most people use about 200 characters. That is like buying a house and only living in the hallway.
Second: the /plan command. This puts Claude Code into planning mode where it maps the full approach before touching code. The source confirms it is a dedicated product feature with its own tools for entering and exiting plan mode. For anything beyond a trivial change, planning first saves tokens and produces better results. It is the difference between a developer who codes immediately and one who thinks first.
Third: permissions. Every "Allow this action?" prompt is a failure of configuration. The source reveals wildcard permission patterns like Bash(git *) that let you approve entire categories of actions once. Configure it properly and Claude Code goes from needing constant approval to operating like an actual agent.
Each of these has a walkthrough in the full guide.
What this means for your business
The gap between how Claude Code is designed and how most people use it is your competitive advantage.
Better models are coming. Anthropic's roadmap includes features like voice mode, daemon mode and a coordinator system that are already flagged in the source code. When those ship, the people who already understand the architecture will be ready to use them immediately. Everyone else will be starting from scratch.
The harness matters as much as the model. If your team is using AI agents in any capacity, the operating environment around those agents determines whether you get 10% or 100% of the value. Claude Code's architecture proves this point - the model is one component in a much larger system.
What to do about it
Three actions, in order of priority:
- Read the power user guide. Nine insights that translate the source code into practical habits. The sections on CLAUDE.md and /plan alone will change how you work.
- Update your installation. Use the native installer, not npm. The leak happened through the npm distribution channel, and a malicious axios package was briefly published around the same time.
- Check your setup. If you installed via npm on 31 March, verify your dependencies. The supply-chain risk window was narrow but real.
Two leaks. 512,000 lines of code. The gap between how this tool is designed and how people actually use it is enormous. Start with the guide and close that gap.
