Anthropic Accidentally Leaked Claude Code's Source—The Internet Is Keeping It Forever

Anthropic didn't mean to open-source Claude Code. But connected Tuesday, the institution efficaciously did—and not adjacent an service of lawyers tin enactment that toothpaste backmost successful the tube.

It started with a azygous file. Claude Code mentation 2.1.88, pushed to the npm registry successful the aboriginal hours of Tuesday morning, shipped with a 59.8MB JavaScript root map—a debug record that tin reconstruct the archetypal codification from its compressed form. These files are generated automatically and are expected to enactment private. But a azygous enactment successful the disregard settings fto it spell retired with the release.

Intern and researcher Chaofan Shou, who appears to beryllium among the archetypal to spot the file, posted a download link to X astir 4:23 a.m. ET, and watched 16 cardinal radical descend connected the thread. Anthropic yanked the npm package, but the net had already archived 512,000 lines of codification crossed 1,900 antithetic files that marque up a large portion of the project.

"Earlier today, a Claude Code merchandise included immoderate interior root code. No delicate lawsuit information oregon credentials were progressive oregon exposed," an Anthropic spokesperson told Decrypt. "This was a merchandise packaging contented caused by quality error, not a information breach. We're rolling retired measures to forestall this from happening again."

The leak exposed the afloat interior architecture of what is arguably 1 of, if not the astir blase AI coding cause connected the market: LLM API orchestration, multi-agent coordination, support logic, OAuth flows, and 44 hidden diagnostic flags covering unreleased functionality.

Among the finds: Kairos, an always-on inheritance daemon that stores representation logs and performs nightly "dreaming" to consolidate knowledge. And Buddy, a Tamagotchi-style AI favored with 18 species, rarity tiers, and stats including debugging, patience, chaos, and wisdom. There’s a teaser rollout for this “Buddy” seemingly planned for April 1-7.

Then there's the item that made everyone connected Hacker News cackle. Per leaker Kuberwastaken, buried wrong the codification was "Undercover Mode"—a full subsystem designed to forestall the AI from accidentally leaking Anthropic's interior codenames and task names erstwhile contributing to open-source repositories. The strategy punctual injected into Claude's discourse virtually says: "Do not stroke your cover."

Apparently, Anthropic began issuing DMCA takedowns against GitHub mirrors. That's erstwhile things got interesting.

A Korean developer named Sigrid Jin—featured successful the Wall Street Journal earlier this period for having consumed 25 cardinal Claude Code tokens—woke up astatine 4 a.m. to the news. He sat down, ported the halfway architecture to Python from scratch utilizing an AI orchestration instrumentality called oh-my-codex, and pushed claw-code earlier sunrise. The repo deed 30,000 GitHub stars faster than immoderate repository successful history.

It’s fundamentally a translation of each the codification from the archetypal connection to Python, truthful technically not the aforesaid thing, right? We’ll permission that to lawyers and tech philosophers.

The ineligible logic present is sharp. Gergely Orosz, laminitis of The Pragmatic Engineer newsletter, argued successful a post connected X: "This is either superb oregon scary: Anthropic accidentally leaked the TS root codification of Claude Code. Repos sharing the root are taken down with DMCA. BUT this repo rewrote the codification utilizing Python, and truthful it violates nary copyright & cannot beryllium taken down!"

It's a clean-room rewrite. A caller originative work. DMCA-proof by design.

The copyright space gets thornier erstwhile considering the ineligible presumption of AI-generated work, and however muddy the criteria gets erstwhile lawyers person to regularisation whether oregon not it carries automatic copyright. The DC Circuit upheld that position successful March 2025, and the Supreme Court declined to perceive the challenge.

If important chunks of Claude Code were written by Claude itself—which Anthropic's ain CEO has implied—then the legal standing of immoderate copyright assertion gets murkier by the day.

Decentralization adds different furniture of permanence. The relationship @gitlawb mirrored the archetypal codification to Gitlawb, a decentralized git platform, with a elemental message: "Will ne'er beryllium taken down." The archetypal remains accessible there. A abstracted repository has compiled all of Claude's interior strategy prompts, which is thing that punctual engineers and jailbreakers volition admit arsenic it gives much insights into the mode Anthropic conditions its models.

This matters beyond the drama. DMCA takedowns enactment against centralized platforms. GitHub complies due to the fact that it has to. Decentralized infrastructure—which powers Gitlawb, torrents, and cryptocurrency itself—doesn't person the aforesaid azygous constituent of failure. When a institution tries to propulsion thing backmost from the internet, the lone question is however galore mirrors beryllium and connected what benignant of infrastructure. The reply here, wrong hours, was: enough.