I built this after deploying OpenClaw on an Ubuntu server.
When the agent needed web search / browsing / scraping, the usual options were either headless browser automation or calling APIs directly. My first instinct was that there should be another path: use a real headed browser in an isolated sandbox, and let a human step in through VNC/Xpra for login or other sensitive steps.
I do the first step when needed, and the agent does the remaining 99+ steps.
And if I don’t fully trust an action, I can still watch or take over.
So I made a self-hosted browser sandbox with:
- remote GUI access via VNC/Xpra
- CDP for Playwright / agent-browser tools
- multiple isolated browser sandboxes for different accounts/login states
- basic file handling for downloads
- simple single-user auth
The main use case is personal/self-hosted agent workflows, not multi-tenant SaaS.
Also if you run agents on your own Mac Mini, maybe you can use this to change another Chrome profile and keep your auth info separate from the agent's.
