A scripted macOS VM (via tart, Apple Virtualization.framework) for running a second, isolated dev environment on the same Mac: Claude desktop + Claude Code, Rust, bun, and screenpipe (app + source). Useful for running two screenpipe instances side by side (separate port 3030, tray, TCC permissions, database), testing onboarding/permission flows from a clean slate, or letting agents work in a sandbox.

./scripts/dev-vm/setup.sh   # one-time: pulls image (~30GB download), creates VM, provisions everything
./scripts/dev-vm/run.sh     # open the VM window (auto-login as admin)
./scripts/dev-vm/ssh.sh     # shell into it

Defaults: screenpipe-dev VM, 8 CPU, 16GB RAM, 80GB sparse disk, macOS Tahoe base image. Display: 1920x1022px virtual display (1:1 crisp in a maximized window on a 1080p host screen; set VM_DISPLAY=2560x1440px or similar on bigger/Retina monitors), window opens maximized (VM_WINDOW=keep to skip). Refit is off on purpose: --display-refit combined with a pinned in-guest mode makes the view render the raw framebuffer 2x-cropped. Other overrides: VM_NAME, IMAGE, VM_CPU, VM_MEM, VM_DISK, GUEST_RES.

• Xcode Command Line Tools, Homebrew, git, gh, jq
• Rust (rustup, stable), bun
• Claude desktop (brew --cask claude) and Claude Code (native installer)
• screenpipe.app (latest stable from screenpi.pe) and a source checkout at ~/Documents/screenpipe

Guest credentials: admin / admin (cirruslabs image default). setup.sh installs an ssh key (~/.ssh/screenpipe-dev-vm) on first run.

1. Open Claude.app and log in (browser OAuth).
2. Run claude once in Terminal to authenticate Claude Code.
3. Open screenpipe.app and grant screen recording, accessibility, microphone. The VM has its own TCC database, so this never touches host permissions.
4. Optional: System Settings > sign in with Apple ID (works because host and guest are both macOS 15+). The App Store does NOT work in any Apple Silicon VM; use direct downloads or brew.

• Max 2 macOS guest VMs running at the same time per host.
• No Apple Neural Engine in guests: Vision OCR falls back to CPU/GPU (works, slower).
• Audio input device support is partial; mic-dependent flows can be flaky.
• GPU is paravirtualized Metal at roughly 90% of host performance.

tart stop screenpipe-dev                      # shut down
tart run screenpipe-dev --no-graphics &       # headless boot
tart clone screenpipe-dev screenpipe-dev-snap # cheap snapshot (APFS clone) while stopped
tart delete screenpipe-dev                    # nuke the VM
tart delete ghcr.io/cirruslabs/macos-tahoe-base:latest  # drop the OCI image cache

Disk reality check: the provisioned VM costs ~35GB of real host disk (grows with cargo target dirs, up to the 80GB sparse cap). The VM is an APFS clone of the image cache, so deleting the cache right after cloning reclaims little; it mainly means a future setup.sh re-clone has to re-download ~30GB.