Show HN: Jev drives a browser – HN login and search in 8s for half a cent

Developer zurfyx published jev-browser-skill, a three-file, dependency-free repository and demo site showing the Jev / OpenAI Decisions API driving a browser by extracting each page's controls into a table and picking an operation and target with a probability for every option. The Hacker News login replay completes in 8 seconds for under half a cent, with Jev averaging 250ms per response (123–495 ms on the HN run) and each stateless step costing a few thousand input tokens, under a tenth of a cent. Jev never sees the DOM and cannot generate text, so any text input must be supplied ahead of time by the user or a complementary LLM.

I put this repo and site together as an educational reference for folks getting started with Jev / OpenAI Decisions API. 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 https://jev-browser.rovira.sh/ hackernews-login , Amazon https://jev-browser.rovira.sh/ amazon-cart 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 https://github.com/zurfyx/jev-browser-skill Comments URL: https://news.ycombinator.com/item?id=50181149 https://news.ycombinator.com/item?id=50181149 Points: 1 Comments: 0