Anthropic Accidentally Reveals Claude Code Source Code

On March 31, a publishing error on npm led to the accidental disclosure of the complete source code of Claude Code, Anthropic's command-line tool for developers. Although this leak does not include the weights of the artificial intelligence models, it nonetheless reveals intriguing details about the company's future projects.

Anthropic, the company behind the high-performing AI model Claude, accidentally made public the source code of its developer tool, Claude Code. This blunder was caused by a simple publishing mistake on the npm platform. Version 2.1.88 of the package "@anthropic-ai/claude-code" contained a .map file, which normally serves as a "map" between a compressed version of the code and its original version. By inadvertently uploading it, Anthropic revealed the complete plans for its interface.

What was leaked is not the brain of the artificial intelligence developed by the company (the infamous model weights, which would weigh hundreds of gigabytes). It is "just" the source code of a command-line tool written in TypeScript, a programming language derived from JavaScript. Despite this, the application's code is filled with industrial secrets.

By sifting through the files available in the archive, it was possible to uncover several areas of work for the startup for 2026. Here are the most notable findings:

• ULTRAPLAN: a system that allows delegating complex exploration tasks to a remote instance of Claude, capable of working autonomously in the background for 30 minutes.

• KAIROS: transforms Claude into a persistent agent with long-term memory, featuring a "nightly dreaming" process to consolidate and sort its logs.

• Coordinator Mode: the tool will be able to clone itself to have multiple "sub-agents" working in parallel, with a "Master Claude" managing the overall orchestration.

• A new language model: the code reveals that Anthropic is testing a language model named Capybara.

• Frustration Telemetry: trackers to measure user annoyance levels, recording frustrating interactions with the AI.

A more playful aspect of this leak is the BUDDY project, a virtual companion inspired by Tamagotchis, set to launch between April 1 and April 7, 2026. This little assistant, which can take up to 18 different forms, will be integrated directly into the command interface and could become a permanent feature as early as May.

This disclosure comes at a delicate time for Anthropic. The company is already at odds with the Pentagon over a disagreement regarding the use of AI for military purposes. Additionally, the startup faced a leak at the end of March revealing Claude Mythos, its upcoming AI model. The accidental publication of the source code sheds light on strategic projects and innovations planned by the company for the coming years.

In addition to these projects, the leak revealed a project named BUDDY, resembling a digital Tamagotchi. This virtual companion, set to launch between April 1 and April 7, 2026, will be randomly generated at its "hatching," based on the user's identifier. It will live directly in the command interface and can take 18 different forms, such as a duck, a capybara, a robot, a ghost, or even an axolotl. Equipped with procedurally generated statistics and small animations, BUDDY will react to the developer's actions. Initially conceived as an April Fool's joke, this project could become a permanent feature starting in May, adding a playful touch that could attract and retain users.