Claude Desktop spawns 1.8 GB Hyper-V VM on every launch, even for chat-only use

Original Article ↗ Hacker News Discussion ↗

VM behavior & disk/RAM impact

  • Claude Desktop starts a Hyper‑V or similar VM on every launch, even for chat‑only use.
  • Reported VM sizes range from ~1.8 GB RAM to 10–13 GB on disk; some macOS users see ~20 GB consumed when Cowork is triggered.
  • Cowork’s VM bundle is installed by default, is not officially removable, and is silently re‑created if deleted or if Cowork is clicked.
  • Users on 256 GB SSDs find this particularly painful; some “remove” it by deleting contents and locking folder permissions.
  • On Windows, disabling Hyper‑V / “Virtual Machine Platform” or running Claude inside a separate VM can avoid or contain it.

Why the VM exists & design tradeoffs

  • The VM powers Claude Cowork, running tools and code in a sandbox and sometimes mounting specified host directories.
  • Some believe eager startup is a performance optimization and a deliberate attempt to reduce friction to using Cowork.
  • Others argue it should be:
    • Opt‑in
    • Started on‑demand
    • Allowed to shut down when idle
    • Exposed via a clear sandbox on/off toggle
  • A minority feels the design is fine if the VM exits with the app; another group is uneasy about an opaque “black box” with unclear observability or logging.

Software quality, “vibecoding,” and UX

  • Many comments criticize Claude’s desktop, CLI, and TUI as slow, buggy, and over‑engineered (React‑heavy, large codebase) relative to a chat client.
  • Users report broken cross‑platform links (e.g., macOS URLs shown on Windows), scrambled terminals, and poor performance in both desktop and web UIs.
  • Several attribute this to “vibe‑coded” features, heavy reliance on LLM‑generated code, and lack of careful human review.
  • Others counter that rapid iteration and shipping partially‑baked features is expected in the current AI race, especially for client “glue” around backends.

Security, sandboxing, and usage patterns

  • Some see per‑app VMs as the right long‑term model: self‑modifying, strongly sandboxed agents with controlled I/O.
  • Others already run Claude/agents with full system access and consider elaborate sandboxing overkill for personal use, relying on backups and source control.
  • There is broad agreement that prompts and local code sent to the model are already going to the provider’s servers; trust mainly comes from policies and contracts, not just the presence of a local VM.

Meta & language

  • Multiple comments nitpick the original title’s double negative (“without no way”), though others view this as unhelpful pedantry.