[BUG] Cowork VM fails to boot on Intel iMac — Claude Desktop burns 120% CPU nonstop

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Cowork local VM never boots on Intel iMac; hypervisor process spins at 120% CPU forever.

Environment: iMac20,1 (27-inch 2020), Intel Core i7-10700K, macOS 26.5.2 (25F84), Claude Desktop 1.32352.1 (auto-updated 2026-08-18 00:14).

Symptom: com.apple.Virtualization.VirtualMachine ("Claude's") holds 120-122% CPU permanently, fan runs loud continuously. Quitting Claude Desktop stops it; relaunching brings it back within minutes.

Cause: every VM startup stalls at guest_vsock_connect started and never reaches completed. The guest kernel boots (reaches Run /init as init process) then hangs; coworkd never starts. The app retries every 2-10 minutes, cannot stop the previous hung VM (Timeout waiting for VM to stop (state: 1)), and spins 4 vCPUs indefinitely.

From cowork_vm_swift.log:

grep -c 'guest_vsock_connect started'   -> 82
grep -c 'guest_vsock_connect completed' -> 61

All 61 successes predate 2026-08-18 16:57. All 16 attempts after that failed.

Likely lead: Direct-boot artifacts not present in bundle; using EFI (GRUB) boot is logged on every attempt, but vmlinuz (14M) and initrd (27M) ARE present in the bundle. The app silently falls back to EFI/GRUB boot, which does not produce a working guest here.

Already ruled out:

• VPN (Tailscale + OpenVPN Connect fully quit, relaunched -> still fails). Note the app's own network diagnostics ran 32 times and reported vmnet unreachable — this is a misleading false signal, same as reported in #24070.
• Corrupt bundle: rm -rf vm_bundles Cache "Code Cache", clean 1.2GB re-download -> same failure.
• Disk space: 216 GB free.
• Thermal throttling: kernel_task stayed at 3-5%.

Not a duplicate of #50627 — that report freezes AFTER the VM boots successfully; here guest_vsock_connect never completes, so the VM never becomes usable at all.

Not a duplicate of #24070 either — same failure mode, but its rm -rf vm_bundles workaround does NOT fix this case (verified: clean 1.2GB re-download, identical failure).

Related: #71557 (Intel Mac regression), #22543, #30972.

What Should Happen?

The Cowork VM should complete startup: guest_vsock_connect should reach "completed", coworkd should start inside the guest, and the workspace should become usable.

Failing that, the app should fail gracefully instead of spinning forever:

1. After N consecutive failed boot attempts, stop retrying. A permanently broken VM should not keep 4 vCPUs at 100% indefinitely — that is the actual user-visible harm (constant full-speed fan with no explanation of why).
2. Surface the real error in the UI ("workspace failed to start, guest did not respond") instead of a misleading VPN warning.
3. Log the guest console output on boot failure so the cause is diagnosable.
4. Provide a user-facing setting to disable the local Cowork workspace/VM entirely, so chat-only users are not affected.

Error Messages/Logs

=== ~/Library/Logs/Claude/cowork_vm_swift.log — the retry loop, repeating indefinitely ===

2026-08-19 00:26:57 [info] startVM called for .../claudevm.bundle with memoryGB=4 cpuCount=auto networkMode=gvisor
2026-08-19 00:27:57 [warn] Timeout waiting for VM to stop (state: 1), proceeding with cleanup
2026-08-19 00:27:57 [info] Direct-boot artifacts not present in bundle; using EFI (GRUB) boot
2026-08-19 00:27:57 [info] Found smol-bin image at /Applications/Claude.app/Contents/Resources/smol-bin.x64.img
2026-08-19 00:27:57 [info] Using gvisor user-mode networking
2026-08-19 00:27:57 [info] Configuration created:
2026-08-19 00:27:57 [info]   - CPUs: 4
2026-08-19 00:27:57 [info]   - Memory: 4GB
2026-08-19 00:27:57 [info]   - Boot: EFI (GRUB)
2026-08-19 00:27:57 [info] Linux VM started successfully
2026-08-19 00:27:57 [info] VM startup step: vm_boot completed
2026-08-19 00:27:57 [info] VM startup step: guest_vsock_connect started
   <- never completes; no further lines for this attempt

2026-08-19 00:35:41 [info] startVM called ...                     (same cycle again)
2026-08-19 00:38:22 [info] startVM called ...                     (and again)
2026-08-19 00:40:23 [info] startVM called ...                     (and again)
2026-08-19 00:46:54 [info] startVM called ...                     (and again)

=== ~/Library/Logs/Claude/coworkd.log — guest kernel boots, then hangs ===

[    0.992261] Run /init as init process
[    1.046432] NET: Registered PF_VSOCK protocol family
[    1.049180] virtio_net virtio0 enp0s1: renamed from eth0
[    1.055043] [drm] Initialized virtio_gpu 0.1.0 0 for 0000:00:0a.0 on minor 0
[    1.659267] tsc: Refined TSC clocksource calibration: 3791.999 MHz
[    1.660617] clocksource: Switched to clocksource tsc
   <- output stops here. coworkd never starts. No further guest output on any attempt.

=== Bundle contents — direct-boot artifacts ARE present despite the log claiming otherwise ===

-rw-------@  128K  efivars.fd
-rw-------@   27M  initrd
-rw-------@   27M  initrd.zst
-rw-------@   10G  rootfs.img
-rw-------@  1.2G  rootfs.img.zst
-rw-r--r--@   58M  sessiondata.img
-rw-------@   14M  vmlinuz
-rw-------@   14M  vmlinuz.zst

=== Misleading network diagnostics (ran 32 times, grew cowork_vm_node.log to 6.9MB) ===

[networkDiagnostics] === vmnet kext status ===
   (empty output)
REACH : flags 0x00000000 (Not Reachable)
No IPv6 states found
   -> These fire even with all VPN software fully quit. Same false VPN signal as #24070.

Steps to Reproduce

1. Intel Mac (iMac20,1, Core i7-10700K), macOS 26.5.2, Claude Desktop auto-updated to 1.32352.1 on 2026-08-18 00:14.
2. Launch Claude Desktop. (No need to open the Cowork tab — the VM starts eagerly at app launch, per #30972.)
3. Open Activity Monitor, sort by % CPU. Observe com.apple.Virtualization.VirtualMachine (shown as "Claude's") climb to and hold 120-122% within a few minutes. Idle Wake Ups ~1780. Fan runs at full speed.
4. Check the VM log:

``
tail -f ~/Library/Logs/Claude/cowork_vm_swift.log
`
Every startup reaches VM startup step: guest_vsock_connect started and stops there. No completed` line ever follows.

1. Confirm the guest hangs:

``
tail -100 ~/Library/Logs/Claude/coworkd.log
`
Kernel boots normally through Run /init as init process and clocksource: Switched to clocksource tsc`, then output stops. coworkd never starts.

1. Count attempts vs successes:

``
grep -c 'guest_vsock_connect started' ~/Library/Logs/Claude/cowork_vm_swift.log -> 82
grep -c 'guest_vsock_connect completed' ~/Library/Logs/Claude/cowork_vm_swift.log -> 61
``
All 61 successes predate 2026-08-18 16:57. All 16 attempts after that failed.

1. Observe the retry loop repeat every 2-10 minutes indefinitely, each cycle logging Timeout waiting for VM to stop (state: 1) — the previous hung VM is never reaped.
2. Cmd-Q Claude Desktop: CPU drops to zero immediately. Relaunch: back to 120% within minutes. Reproducible every time.

Attempted fix that did NOT work:

# Cmd-Q Claude Desktop first, then:
rm -rf ~/Library/Application\ Support/Claude/vm_bundles \
       ~/Library/Application\ Support/Claude/Cache \
       ~/Library/Application\ Support/Claude/Code\ Cache

Relaunch -> clean 1.2GB download, re-extracted to 10GB rootfs.img -> identical failure.

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Unknown — the Claude Desktop build in use before the 2026-08-18 00:14 auto-update (version number not recorded; VM booted fine until 2026-08-18 16:57)

Claude Code Version

Claude Desktop 1.32352.1 (this is a Claude Desktop / Cowork VM bug, not Claude Code CLI)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Hardware/OS

| Item | Value |
|---|---|
| Machine | iMac20,1 (27-inch, 2020) |
| CPU | Intel Core i7-10700K @ 3.80GHz |
| macOS | 26.5.2 (25F84) |
| Claude Desktop | 1.32352.1 (auto-updated 2026-08-18 00:14) |
| VM bundle | re-downloaded 2026-08-19 00:30 (clean) |
| VM config | CPUs 4, Memory 4GB, networkMode=gvisor, Boot: EFI (GRUB) |
| Guest image | smol-bin.x64.img (x86-64 guest) |
| Free disk | 216 GB |

Impact

The machine's fan runs at full speed continuously for as long as Claude Desktop is open, with no indication in the UI of why. The only workaround is to quit the app entirely (Cmd-Q); closing the window is not enough, since the VM starts eagerly at app launch (#30972) and there is no setting to disable it.

Note that this is not limited to users who use Cowork — the VM starts even if the Cowork tab is never opened.

Timeline

| Time | Event |
|---|---|
| 2026-08-18 00:14 | Claude.app auto-updated to 1.32352.1 |
| 2026-08-18 16:08 | new VM bundle downloaded |
| 2026-08-18 16:57 | first failure; every attempt from here on fails (0/16) |
| 2026-08-19 00:30 | manual rm -rf vm_bundles + clean re-download |
| 2026-08-19 00:35+ | still fails identically |
| 2026-08-19 (next day) | still failing, unchanged, app still 1.32352.1 |

Full logs available on request.