Letting an agent run commands on your machine is nerve-wracking. Give it a box instead: a real, isolated Linux environment it can fully take over, that you can throw away when it’s done. 
zomg create agent-run
zomg exec agent-run -- bash -lc 'whatever the agent wants to run'
zomg delete agent-run --force
Nothing it does touches your laptop. Delete the box and it’s all gone.

​
Branch an experiment with a box

Want to try two approaches from the same starting point? Box an existing box and you get a full copy of its filesystem in seconds — cheap, because it’s a copy-on-write snapshot: 
zomg create idea-a --fork agent-run
zomg create idea-b --fork agent-run
Each box runs independently. Keep the one that works, delete the rest.

​
Snapshot a good state

Capture a box you can return to later, or turn it into a reusable starting point for new boxes: 
zomg snapshot create agent-run good-state

​
Keep the agent signed in

Coding agents need their own credentials inside the box. Set up a developer profile once and Zomg syncs Claude Code and Codex logins into every box in the project — no copying tokens around each time. See Coding agent auth.

​
See what the agent did

Boxes export OpenTelemetry traces for Claude Code and Codex, so you can review an agent’s run after the fact: 
zomg observability traces agent-run --agent claude --since 1h

​
Keep going

Snapshots and bases

Capture box state and turn it into a reusable base.

Coding agent auth

Keep Claude Code and Codex authenticated inside boxes.

Boxes

The box lifecycle, including boxes.

Development

Run agents over SSH and review their traces.