Jev Ultrafast: A Browser Agent That Never Lets the Model Write a Selector
It's called Jev Ultrafast. Give it a URL and a goal in plain English, and it clicks, types and selects its way through the site.
Every decision comes from Jev, TypeSafe's new decision model. It takes the current state and a list of options and picks one (click this, type that, yes/no) in a few hundred milliseconds, with probabilities your code can actually use.
Here's how that plays out on a one-way flight search:
1️⃣ The agent reads the page and numbers every visible control you can interact with:
[1] button · Round trip
[2] combobox · Where from?
[3] combobox · Where to?
2️⃣ Jev reads that list and the goal, and answers with one action and one number. For example, CLICK [1] to change the trip type, or TYPE_TEXT [2] for the origin.
3️⃣ If the action is typing, a small LLM (Mercury 2.5) writes the text: "Zurich".
4️⃣ The agent acts, reads the page again, and repeats until Jev says DONE.
Most browser agents send a screenshot to a big vision model at every step. Jev gets a short text list and makes one pick, with a median decision time of 178 ms.
Jev only ever picks an operation and a number from the list. Nothing a model outputs becomes a selector or code, and the agent rechecks the element before acting.
Result: the full Google Flights search took 7.1 seconds in real time, including Google's loading, and cost $0.0039.
That makes it practical for search on real sites, filling out forms, working with sites that have no API, and prototyping agents cheap enough to run often.
Jev Ultrafast hit 14k+ stars on GitHub in just 5 days.
Jev Ultrafast, built by Browser Use on top of TypeSafe's Jev decision model, does multi-step web tasks from a URL and a plain-English goal. It numbers every interactive element on the page, sends that list plus the goal to Jev, and gets back one operation and one target — click, type, select, scroll, or done — rather than a screenshot going to a large vision model at every step. A small text model only gets involved to generate actual typed input.
Because the model only ever picks from a fixed list of real, observed page elements — never emitting a selector or code — the agent can recheck the element before acting, which is the project's main safety property.
Browser Use's own demo reports a one-way Google Flights search completing in 7.1 seconds including page load, at a cost of a fraction of a cent — though that figure comes from the project itself and hasn't been independently verified yet. The repo, created September 16, 2026, has grown quickly, passing 10,000+ stars within its first few days.
Worth knowing before relying on it for anything unattended: independent analysis of Jev-style decision models has flagged real architectural limits, including p99 latency well above the typical case and a documented tendency to favor earlier-listed options over later ones with equivalent meaning — a bias worth testing for specifically in an app built around numbered lists of page elements.