🔥 Anthropic accidentally leaks its own source code, revealing unreleased "Capybara" model

Key Points

• Anthropic inadvertently exposed the full internal source code for Claude Code, its viral AI coding assistant, via a public registry.

• The leak revealed the "agentic harness" and unreleased model tiers known as Mythos and Capybara.

• Within hours, a community-driven repository named ClawCode reached 50,000 GitHub stars, shattering records set by previous open-source alternatives like OpenClaw.

• An Anthropic spokesperson stated that "no sensitive customer data or credentials" were involved in what they called a "release packaging issue."

Anthropic has accidentally leaked the internal source code for its popular AI coding assistant, Claude Code, according to the company and multiple reports on Tuesday.

The leak occurred when a 60MB source map file was inadvertently included in the version 2.1.88 release of the @anthropic-ai/claude-code package on NPM, a public registry for JavaScript developers. The file effectively acted as a "blueprint," allowing developers to recompile roughly 512,000 lines of TypeScript code and 1,900 internal files into a human-readable format.

While Anthropic confirmed the authenticity of the leak to Gizmodo, a spokesperson emphasized that the incident was caused by "human error" during the release process rather than a security breach.

The "Agentic Harness" and "Capybara" Exposed

According to Fortune, the leak is particularly damaging because it exposes the "agentic harness", the sophisticated software layer that allows the Claude model to interact with terminal commands, edit files, and execute autonomous coding tasks.

Beyond the tool’s mechanics, the code provided concrete evidence of Anthropic’s secret model roadmap. References to a new tier called "Capybara" (also referred to as Mythos) were found, suggesting a model that sits above the current flagship, Claude Opus 4.6, in both cost and capability.

The Rise of "ClawCode"

The developer community reacted with unprecedented speed. Shortly after the leak was publicized by security researcher Chaofan Shou on X, a mirror repository named ClawCode (hosted at instructkr/claw-code) appeared on GitHub.

According to sources tracking the repository, ClawCode amassed 50,000 stars in just two hours, becoming the fastest-growing repository in the platform's history. It quickly surpassed the momentum of OpenClaw, a previous community effort to replicate Anthropic's tool, as developers rushed to fork the raw, leaked logic before Anthropic could issue DMCA takedowns.

Strategic Fallout

The exposure of Claude Code’s internal logic is a major blow to Anthropic's competitive "moat." With a run-rate revenue of $2.5 billion as of February, the tool is a core pillar of Anthropic’s business. Competitors and open-source developers can now analyze the company’s internal orchestration strategies, memory management, and tool-invocation loops.

"This is like giving away the kitchen blueprints while you're still the most popular restaurant in town," says one industry analyst. "The underlying model weights are still safe, but the 'secret sauce' of how that model actually operates in a real developer's terminal is now public domain."

Personal Take:

Honestly, this is the "Vibe Coding" era's biggest facepalm. Anthropic spends millions on AI safety, yet they got taken down by a .map file. By leaking the agentic harness, they’ve effectively "forced" an open-source standard for AI agents. Whether you’re team Anthropic or team OpenAI, the barrier to building a world-class coding bot just hit zero. Grab your popcorn; the next 48 hours on GitHub will be historic.