/ runtime·vm-7f3a · us-west
live
agent / taskresources
- clauderunning
refactor auth middleware
4 vCPU16 GB - codexrunning
write integration tests
8 vCPU32 GB - grokqueued
benchmark cold starts
4 vCPU16 GB
→ provisioned in 1.4s · isolated rootfs · persistent workspace
OpenAI
Claude
Grok
Gemini
Cursor
1.2s
avg. VM spin-up
0
workflow changes
24/7
agent loops
- $ mm market search agent-nodenode-agent-22.04 4 vCPU · 16GB · readyclaude-devbox 8 vCPU · 32GB · ready$ mm market launch node-agent-22.04→ vm-9c2e provisioned in 1.4s
- $ mm license scan ./src --diff HEAD~1⚠ GPL-3.0 lodash-fork@1.2.0 (agent-added)✓ MIT 48 packages clean→ blocked merge until license reviewed
- $ mm deploy --env prod --from vm-7f3abuilding image… done (41s)rolling 3 / 3 healthy→ https://api.northline.dev
- $ mm debug attach vm-7f3asession claude#14 pausedauth.ts:88 TypeError: token undefined→ attached · live stdout streaming
- $ mm share vm-7f3a --with elena@silt.devinvite sent · shell + workspace→ elena joined · 2 sessions active
“We used to leave Claude overnight on a spare Mac Mini and hope it didn't melt. Moving the same workflow onto a minimachines VM cut our iteration loop roughly in half — and the laptop fans finally shut up.”
“The copyleft checker caught a GPL transitive dep an agent had pulled mid-refactor. That alone paid for the pilot. Cloud debug is what we actually live in day to day though — attaching to a stuck session beats SSH archaeology.”
“Collaborative cloud is the underrated part. I can jump onto the same VM my teammate's Codex run is using, poke the failing test, and hand it back without shipping zip files around Slack.”