I built Slopera, a desktop browser that, instead of connecting to the real web, renders each page as what an LLM thinks it might look like. You'll never get a 404, basically any URL you or the LLM (for example, in fake google results) comes up with exists. Clicking a link creates the next page on demand.

Images and interactive JS content are generated too.

It's Electron + TypeScript. You need an Anthropic or OpenRouter key to run it. Support for local models is planned, which would make this a fully offline browser.

The name is a pun on Opera / slop / era.

One thing I didn't expect: the models differ a lot in how "faithful" they get when you type a real URL. Of course smaller models are further from the truth, but it seems to me that the big models from OpenAI and Anthropic tend to try to prevent copying a branded web site, while the Chinese model GLM 5.2 often delivers almost 1:1 copies of the layout of sites like wikipedia.org etc. Feels like a training/RLHF difference around reproducing recognizable content. Does anyone have more insight into that?