A story about leaked Claude Code source files recently went viral. One npm package version accidentally included a large source map that exposed the TypeScript source of the client application.
An important distinction: this was not a leak of Claude’s model weights or the source of the language model itself. It is still useful material for the local-LLM tooling ecosystem because it offers a closer look at the agent client, its tools, and its engineering decisions.
For Anthropic, this was clearly an incident rather than a healthy way to distribute code. At the same time, it highlights the broader centralization problem once again: when compute, interfaces, and access rules are concentrated in a handful of enormous companies, users become more dependent on their decisions, censorship, and product policy.
Other people’s mistakes sometimes produce an unintended public benefit. But a sustainable alternative should not depend on leaks; it should come from open weights, reproducible tools, and the ability to run them on infrastructure you control.
Original Telegram post · incident discussion in the Claude Code repository