Cowork VM daemon fails to auto-launch — "VM is not running. Start a conversation first."

[Echo9k]

Hi there,
First off, thanks for this project — really appreciate the work that makes Claude Desktop possible on Linux.

Description

Cowork mode (both "Code" and "Tasks" sidebar tabs) fails on launch with repeated errors:

Error: VM is not running. Start a conversation first.

The cowork-vm-service.js daemon is not being auto-launched by the Electron app, so the Unix domain socket at $XDG_RUNTIME_DIR/cowork-vm-service.sock never gets created. The Cowork UI loads but cannot communicate with the backend.

Analysis

The auto-launch code in the patched index.js (from PR #234) uses child_process.fork() inside the Ma() retry loop:

process.platform==="linux"&&!Ma._svcLaunched&&(Ma._svcLaunched=true,(()=>{
  try{
    const _d=require("path").join(process.resourcesPath,"app.asar.unpacked","cowork-vm-service.js");
    if(require("fs").existsSync(_d)){
      const _c=require("child_process").fork(_d,[],{detached:true,stdio:"ignore",
        env:{...process.env,ELECTRON_RUN_AS_NODE:"1"}});
      _c.unref()
    }
  } catch(e){}
})())

This fork only triggers when Ma() gets an ENOENT on socket connect. However, the CustomPlugins_listAvailablePlugins handler calls b_t() which checks VM status via a different path — it throws "VM is not running" before Ma() ever gets a chance to auto-launch the daemon. The catch(e){} also silently swallows any fork errors.

Starting the daemon manually works perfectly:

node /usr/lib/claude-desktop/node_modules/electron/dist/resources/app.asar.unpacked/cowork-vm-service.js
# [cowork-vm-service] Service started on /run/user/1000/cowork-vm-service.sock

Once the socket exists, both "Code" and "Tasks" tabs function correctly.

Workaround

A systemd user service reliably solves the problem:

# ~/.config/systemd/user/cowork-vm-service.service
[Unit]
Description=Claude Desktop Cowork VM Service Daemon
Documentation=https://github.com/aaddrick/claude-desktop-debian

[Service]
Type=simple
ExecStart=/usr/bin/node /usr/lib/claude-desktop/node_modules/electron/dist/resources/app.asar.unpacked/cowork-vm-service.js
Restart=on-failure
RestartSec=3
Environment=COWORK_VM_DEBUG=1

[Install]
WantedBy=default.target

systemctl --user daemon-reload
systemctl --user enable --now cowork-vm-service.service

Environment

• Claude Desktop: 1.1.3189-1.3.11 (.deb, amd64)
• Linux: Ubuntu x86_64, kernel 6.17.0
• Claude Code CLI: 2.1.45
• Node.js: system /usr/bin/node

Steps to Reproduce

1. Install latest claude-desktop .deb from this repo
2. Open Claude Desktop
3. Click "Code" or "Tasks" in the sidebar
4. Observe: UI loads but sessions fail with "VM is not running"

Verification

# Confirm the daemon is NOT running
ss -lx | grep cowork
# -> (empty)

# Check logs for the error
grep "VM is not running" ~/.cache/claude-desktop-debian/launcher.log
# -> Error occurred in handler for '..._CustomPlugins_$_listAvailablePlugins': Error: VM is not running. Start a conversation first.

# Start daemon manually — Cowork immediately works
node /usr/lib/claude-desktop/node_modules/electron/dist/resources/app.asar.unpacked/cowork-vm-service.js &
ls $XDG_RUNTIME_DIR/cowork-vm-service.sock
# -> /run/user/1000/cowork-vm-service.sock

Suggested Fix

Ship a systemd user service with the .deb package. This is more reliable than the in-app fork approach because:

• The daemon starts before Claude Desktop, so the socket is ready immediately
• Survives app restarts without race conditions
• Auto-restarts on failure
• Standard Linux service management (systemctl --user status/restart/stop)

[Echo9k]

Opened PR #237 with a fix for this — ships a systemd user service with the .deb package so the daemon starts reliably on login, bypassing the in-app fork() race condition.

---

Written by Claude Opus 4.6 via Claude Code

[aaddrick]

Hey Echo9k! Great diagnosis and thanks for the PR.

I looked at #237 and found two gaps, so I opened a PR against your branch with fixes on top:

Echo9k#1

The two things I addressed:

1. The systemd service was calling /usr/bin/node directly, which isn't a declared package dependency. Switched it to use the bundled Electron binary with ELECTRON_RUN_AS_NODE=1 instead.
2. AppImage users had no coverage from the systemd approach. Added a pre-launch block in the AppRun script that starts the daemon before Electron loads.

Worth merging those before this lands in main. Happy to adjust anything if the approach doesn't fit your intent.

---

Written by Claude Sonnet 4.6 via Claude Code