dejevu

Checked, not fully reviewed. We ran its 48 tests (all pass) and drove the --backend typesafe policy against a stand-in. We did not run Chrome or any model, so the timings below are the author’s.

dejevu reads the page into a numbered list of controls and asks a model for one action at a time from a fixed menu — click, type, select, press, scroll, wait, done, blocked. The default backend is any OpenAI-compatible endpoint, OpenRouter out of the box; Jev is not needed. Every task ends with a check in code on the final page, not the model’s word that it is done, and the README adds three hold-out tasks written after tuning. The author reports llama-3.3-70b on Groq finishing a nine-step Google Flights search 3 of 3 times at a median of 5.6 s, and smaller models stopping inside the calendar.

--backend typesafe is the comparison arm: one System One request with an operation choice and a target choice per operation, plus a text model to write field values. Against our stand-in it sent the published request shape and read answers keyed by question name; on a 529 it retried and then clicked; three 529s in a row stopped it with no action; and an operation that was not offered was refused with “no action executed”. The retry is a fixed short backoff that ignores retry-after.

The README’s head-to-head sets its own runs beside the published jev-ultrafast demo — a different date, machine and network, and not the TypeSafe backend this repository ships. The README itself says two runs per cell is “evidence of a difference, not a benchmark”. We do not repeat the Jev figures; to compare, run both backends here on the same machine.

MIT, 11 stars. Worth reading for the action menu and the verification, whichever model you put behind it.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Oct 5, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.