This demonstrates how Jev can drive a browser (the idea being extract it into a skills file). On every step the page's controls are extracted into a table, and Jev picks an operation and a target, with a probability for every option.
Try the HN login replay (https://jev-browser.rovira.sh/#hackernews-login), Amazon (https://jev-browser.rovira.sh/#amazon-cart), or run it live with your own key.
Interesting bits: - Jev mean response time is 250ms (123–495 ms on the HN run). - Jev never sees the DOM, only a table of controls extracted from it ahead of time. It's in some ways, similar to the accessibility tree that good LLM browser skills use now. - Jev can't generate text (in a good way) so this need to be provided ahead of time, either by you or a complementary LLM. - Each step is stateless and costs a few thousand input tokens, under a tenth of a cent.
Repo (three files, no dependencies): https://github.com/zurfyx/jev-browser-skill