BREAKING: Claude Code source leaked

Claude Code’s bundled JavaScript shipped with source maps that effectively exposed large portions of its underlying TypeScript source—triggering a wave of takedowns, speculation, and rapid community reverse-engineering. The leak matters because source maps are designed to map minified production code back to the original code for debugging; when they include the original sources, they turn “obscured” software into something far more inspectable. In this case, the source maps appear to have contained enough material to reconstruct major parts of Claude Code’s codebase, including unreleased features and internal mechanisms.

The chain of events starts with how modern web/Node tooling works: browsers and JavaScript runtimes can’t execute TypeScript directly, so build pipelines transform TypeScript into minified or obfuscated JavaScript. That transformation makes debugging difficult, so developers generate source maps—files that link positions in the shipped code back to the original source. Error monitoring services like Sentry can use those maps to produce readable stack traces without exposing the original code to end users. But for Claude Code, the source maps weren’t kept hidden; the npm package appears to have included a source map folder containing the original source content. The result is that anyone downloading the package could inspect the “real” code behind the obfuscated CLI.

Anthropic has reportedly sent extensive DMCA requests in response to earlier leaks of Claude Code source via the same mechanism, and this incident follows that pattern. The transcript argues the most likely cause wasn’t a deliberate “intentional leak,” but a packaging mistake—possibly tied to efforts to improve production logging after rate-limit issues surged. A key clue is that the leak included references to an internal dependency (Claude Agent SDK 0.2.88) that later disappeared when the release was pulled, breaking local installs for the investigator and forcing downgrades.

A set of conspiracy theories then gets addressed. One claims the leak was intentional; the counterpoint offered is that the original downloadable artifact was taken down quickly and that Anthropic’s history of DMCA enforcement makes “controlled release” unlikely. Another theory blames bun’s web hosting behavior (bun serve sometimes exposes source maps), but the transcript says Claude Code doesn’t use bunserve, and an employee associated with bun reportedly pointed to that mismatch. A further theory suggests competitors can now improve their own harnesses; the transcript counters that Claude Code’s harness performance is comparatively weak versus alternatives and that its source isn’t a magic advantage.

Once the source was reconstructed, the most consequential takeaways are the internal feature flags and unreleased capabilities allegedly embedded in the code. Among the highlighted items are “dream mode” (background agents consolidating memories), “coordinator mode” (parallel worker agents with shared tool access), “ultra plan” and “ultra review” (long-horizon planning and automated review with billing controls), plus “auto mode” and an always-on proactive system referred to as “Chyros.” The leak also surfaces operational details like permission cascades, prompt-cache sharing strategies to reduce inference costs, and extensive feature-flag usage via GrowthBook.

Finally, the transcript pivots to what Anthropic should do next: open-source Claude Code or at least publish a credible timeline, stop aggressive DMCA campaigns against forks and non-infringing activity, and communicate transparently with the community rather than relying on legal suppression. The core message is that the “secret sauce” argument no longer holds once the code is out, and that human, engineer-led updates could turn a PR disaster into a credibility win.