Overview & result
ZephyrDeng shows ego-jev, built on Jev: One ~0.4 s typed decision per DOM step, instead of an LLM turn.
From the original post
ego-jev by @zephyrdeng: One ~0.4 s typed decision per DOM step, instead of an LLM turn.
How Jev fits in the loop
- ego takes a snapshot of the page and numbers the interactive elements.
- One System One call picks the operation and its target in the same request.
- ego clicks, fills or selects in your real browser, with your sessions and logins.
- Free text, screenshots, login, payment and canvas steps go back to the calling agent.
How to reproduce
- Review the source documentation and repository at https://github.com/ZephyrDeng/ego-jev.
- Configure environment with TypeSafe Jev / System One API keys.
- Run test prompts against the decision endpoints.
Why this build matters
Shows Jev in practical use for browser and autonomous agents.
Limitations
- Jev does not type free text, read screenshots, or touch login, payment or canvas.
- Timings (about 300–550 ms and about $0.0005 per step) are the author's.
- Request recording is opt-in and off by default.