I use a cloudtop for my development. I can provision a new cloudtop and download a clean copy of my repo in around an hour. I SSH or remote desktop from my local workstation to that machine for anything I need to run. The agent runs in a web service on that machine, which is how I interact with it.

Cloudtops only have access to credentials for my personal instances. Instances can be wiped and brought back up in about half an hour from a script. Any deployment to staging/prod is done with other tooling that doesn't run through AI, and only looks at checked-in code.

If my AI went completely haywire, it could hose my personal instances and a single cloudtop. It would take about 15 minutes to de-provision the old cloudtop and provision a new one. After that I'd kick off 45 minutes of git cloning, and in parallel run the scripts to wipe my personal instances and reset them to good. After an hour I'd be back to a clean workable state.

I've had to do this twice, but neither time was due to AI, both were due to bad updates my cloudtops got.

I use the Claude Code cloud sessions feature which is similar, it creates a cloud VM and runs all the code on that. It's pretty handy since you can just ask for a change to a repo on the mobile app and it can run everything on the VM and make a PR.

Sometimes I just want to work on local files though, and with Claude code there isn't really a good way to sandbox it while also having access to claude-in-chrome.

I use droneship, it basically feels the same as running locally (droneship run claude) but it runs in a cloud sandbox with easy egress control. It takes running snapshots that you can rollback and you can diff the workspace and pull to bring the changes back to your local.