A massive source code leak tied to Anthropic has offered a rare glimpse into the future of its AI ecosystem. The leak reveals experimental features, hidden tools, and ambitious plans surrounding its Claude Code platform.
The leaked codebase of Claude Code spans more than 512,000 lines across over 2,000 files.
Developers analyzing the data discovered several disabled or inactive features, offering insights into potential upcoming capabilities of Anthropic’s AI systems.
One of the most notable discoveries is “Kairos,” a persistent daemon designed to run in the background—even when the Claude Code terminal is closed.
The system would rely on periodic “tick” prompts to check for tasks and a “PROACTIVE” flag to surface important insights without user prompts.
Kairos also uses a file-based memory system, allowing it to retain context across sessions and better understand user preferences and workflows.
AutoDream: AI that ‘reflects’ on its memory
To manage long-term memory, the code references a system called AutoDream.
This feature activates when users go idle or end a session, prompting the AI to perform a “dream”—a reflective process that reviews and organizes stored information.
The system scans conversations for valuable insights, removes duplicates, resolves contradictions, and updates outdated data to maintain structured memory.
Another feature, labeled “Undercover mode,” appears designed to let AI agents contribute to public open-source repositories anonymously.
The prompts instruct the system to avoid revealing internal details, including model names or any indication that the contributor is an AI.
It also explicitly prevents attribution, raising questions about transparency in AI-generated code contributions.
Meet ‘buddy’: Playful virtual assistant
On a lighter note, the leak reveals “Buddy,” a Clippy-like assistant that interacts with users through small speech bubbles.
Buddy is designed as a “watcher” near the input box, appearing in 18 different forms—ranging from blobs to axolotls—rendered in ASCII-style animations.
According to comments in the code, Buddy was planned for a teaser launch between April 1 and 7, followed by a full release in May, though its future remains uncertain after the leak.
Other features in development
The leaked code also references several additional tools and capabilities:
UltraPlan
An advanced planning feature that allows high-level AI models to create detailed plans, running for 10 to 30 minutes for complex tasks.
Voice Mode
A feature enabling users to interact with Claude Code through voice, similar to other conversational AI platforms.
Bridge Mode
An extension of Anthropic’s Dispatch tool, allowing remote sessions controlled via browser or mobile devices.
Coordinator Tool
A system designed to manage multiple AI workers simultaneously, enabling parallel software engineering tasks through WebSocket communication.
