Rendering web pages is hard, so I had an image generation model render them and made it act like a browser.
The omnibox sends the given text to GPT 5.4 which does it's best to find and download the right resources and describes how they should look to GPT Image 1.5, which draws the page. When you click somewhere on the page, a red circle is drawn around the point you clicked, and that image is given as input to the GPT 5.4 session and it repeats.
An example of rendering my homepage with the annotation for clicking a link: 
It just fakes the youtube homepage with garbage. I didn't check the session logs, but I assume since I don't pretend to be a human user at all it's blocked from seeing what's actually on the homepage. That happened to wikipedia, but instead of hallucinating it just showed the error. 
And of course, an attempt at rendering the hackernews homepage. This was was reasonably accurate at the time, but most of the numbers were completely made up. 