Skip to main content
You want an agent to open web pages, read them and act on them, in a browser that is not yours: no cookies of yours, no access to your network, and gone when the task ends. Pieces. The desktop template has a browser, XFCE, xdotool and scrot. The anona-compute[desktop] Python extra drives it with E2B’s desktop SDK. Live view’s dom channel replays the open page for you, sharp at any size and with passwords masked.
1

Open a page and look at it

browse.py
2

Loop: look, decide, act

Send the screenshot to your model and carry out what it answers with left_click, write, press and the other actions in Desktop. Take a new screenshot after each step.
3

Fence it in

Browsing agents follow links they should not. Allow only the sites the task needs with network={"deny_out": ["0.0.0.0/0"], "allow_out": [...]} (Networking), and use the credential vault for any API token rather than typing it into a page.
The free plan has no public egress for now, so its sandboxes cannot load public pages unless their hosts are in allow_out. See Limits.
Headless instead. For scraping without a screen, install Playwright or Puppeteer in a python or node sandbox, or bake it into your own template so every sandbox starts with it. Runnable example: examples/python/browse_web.py in the public SDK repository. It prints what happens at each step.