Claude Code's Cowork Feature Silently Creates 10GB VM Bundle on macOS
Original: Anthropic Cowork feature creates 10GB VM bundle on macOS without warning View original →
Overview
A GitHub issue revealed that Claude Code's new Cowork feature — designed for multi-agent AI collaboration — silently creates approximately 10GB of virtual machine storage on macOS without warning users. The issue quickly gained traction on Hacker News with over 200 upvotes.
Technical Details
The Cowork feature uses Apple's Virtualization Framework to run agents in isolated sandboxed environments. While this approach has legitimate security benefits, the implementation creates a large VM bundle without user consent or disclosure.
Users running Claude Code inside existing VM environments discovered that Cowork is completely non-functional in nested virtualization scenarios, as Apple's framework doesn't support nested VMs. This effectively blocks Cowork for a large segment of developers using virtualized development environments.
Community Discussion
Hacker News commenters expressed frustration at the broader trend of applications silently consuming large amounts of disk space, drawing comparisons to Apple Podcasts downloading 120GB of episodes without user knowledge. Some commenters also noted skepticism about whether the GitHub issue itself might be AI-generated, given it was explicitly filed "via Claude Code."
Outlook
Anthropic has not yet issued an official response. While VM-based isolation is a valid security architecture for agentic code execution, the community consensus is that upfront disclosure and user consent are necessary before consuming significant system resources.
Related Articles
HN focused less on telemetry as an idea and more on whether opt-out controls work when gh runs inside CI, servers, and automation.
HN’s GitHub CLI telemetry thread turned into a developer-tools trust debate: not whether metrics can help, but whether default-on collection belongs in a command-line tool.
Why it matters: persistent memory is one of the missing pieces between demo agents and useful long-running agents. Anthropic pushed the feature into public beta on April 23 and framed it as a memory layer that learns from every session.
Comments (0)
No comments yet. Be the first to comment!