Anthropic's Claude Code Source Code Leaked, Sparking Rapid Innovation in AI Agents

Fazt
#anthropic #claude-code #ai-agents #source-code-leak #open-source-ai

Anthropic, the developer behind the Claude Code AI assistant, is facing significant repercussions after the accidental leak of its tool’s source code. A security researcher, Shaw, discovered the full codebase of Claude Code version 2.88, which was inadvertently made public due to the inclusion of a large .map file in a published client application. This oversight allowed the source to be downloaded from an R2 bucket and subsequently published on GitHub, despite Anthropic’s rapid attempts to contain the spread. The leak has immediately catalyzed the developer community, leading to numerous re-implementations and forks. Notably, one project, a re-write of Claude Code using OpenAI’s Codex, achieved over 50,000 stars on GitHub within two hours, setting a new record for repository traction. Other projects, like ‘OpenClode,’ have emerged, offering multi-model support for GPT, Llama, Gemini, and DeepSeek, expanding beyond Anthropic’s ecosystem.

The exposed source code offers an unprecedented look into Claude Code’s internal workings, revealing its architecture built on the ink framework—a console-based counterpart to React—and an extensive use of hardcoded prompts for behavior control, safety filtering, and task execution. Developers have identified internal utilities like tool and Bash Tool, implemented in TypeScript, which manage file operations and other core functions. Furthermore, the leak has brought to light several unreleased or rumored features, including the Cairos model, an event-driven AI designed to operate without explicit prompts, and Ultraplan, a planned service for extended remote Claude-based planning sessions. An ‘Undercover’ mode prompt, apparently used internally by Anthropic, was also discovered, designed to mask AI authorship in public code contributions. This event is poised to significantly influence the AI agent landscape, enabling competitors to analyze and integrate Claude Code’s methodologies, ultimately driving innovation and potentially democratizing advanced AI coding assistant capabilities.