Jev for Browser and Computer Use Agents
Last checked · Independent guide, not affiliated with TypeSafe AI
Jev is used in browser and computer-use agents to pick the next action and the element to act on, as a Choice over the options on screen, while a small LLM writes text only when a field needs typing. Browser Use's open-source Jev Ultrafast agent and the typesafe-computer-use project both work this way and report steps costing a fraction of a cent.
Agents that click through websites or desktop apps spend most of their time on small decisions: which button next, which field to fill, whether the goal is done. Frontier models handle these well but slowly, taking several seconds and a screenshot per step. Jev’s pitch is to make each of those decisions in a few hundred milliseconds for a fraction of a cent. The most visible Jev projects in its first week were agents built exactly this way.
Jev Ultrafast by Browser Use
Section titled “Jev Ultrafast by Browser Use”Jev Ultrafast is an open-source (MIT) browser agent published by Browser Use, the company behind the popular browser-use library, on September 16, 2026. It had more than 6,000 GitHub stars three days later.
How it works, according to its README:
- Each observation of the page is turned into a numbered element table: buttons, text boxes and dropdowns with their labels and current values.
- One Jev request picks the operation (
CLICK,TYPE_TEXT,SELECT,SCROLL_UP,SCROLL_DOWN,WAIT,DONEorBLOCKED) and, in the same call, speculative target questions for each operation: which element to click, which to type into, which option to select. - Code uses only the target that matches the chosen operation. Two decisions, one network round trip.
- A small LLM is called only when the operation is
TYPE_TEXT, to write the text itself.
The headline demo searches Google Flights from Zürich to London in 7.1 seconds end to end, including page loads and text generation. Running it requires a TYPESAFE_API_KEY and a key for the text model. Browser Use also advertises a waitlist for a hosted cloud version.
typesafe-computer-use
Section titled “typesafe-computer-use”typesafe-computer-use applies the same idea to a whole Mac. It reads the screen with OCR instead of sending screenshots to a large model, asks Jev a Choice over the possible next actions, and calls a writing model only when a text field needs free text. It requires macOS 14 or newer and Python 3.12.
The project’s README compares one decision against Claude Opus 5 working from a bare screenshot:
| Per decision (project’s own measurement) | Jev | Opus 5 |
|---|---|---|
| Cost | $0.0002 | $0.032 |
| Model latency | 0.13 to 0.38 s | 5.2 s |
| End-to-end step with capture and OCR | about 1.5 s | about 5.5 s |
The same README is candid about the trade-off: the large model read event dates straight from the pixels and compared them unaided, while the Jev version needed date parsing added in code. That is a general rule for Jev agents, covered below.
The pattern behind both
Section titled “The pattern behind both”- Turn the screen into text. Jev accepts text only, so the agent needs an element table, accessibility tree or OCR output as its state.
- Make each step a Choice. Options are the actions and elements currently available. A Choice accepts up to 255 options; TypeSafe’s launch demo of Wikipedia racing handled pages with more links in two stages, scoring candidates first and then choosing.
- Ask speculative questions together. Asking for every possible target alongside the operation costs a few extra tokens and saves a second round trip. TypeSafe documents this as its speculative fan-out pattern.
- Gate on confidence. When the Choice confidence is low, stop, ask the user or hand the step to a bigger model.
- Keep generation separate. Jev cannot write the text to type; a small LLM does that.
Where it breaks down
Section titled “Where it breaks down”TypeSafe’s list of Jev 1.13 weak spots matters more for agents than for most uses:
- Dates and numbers. Comparing prices, dates or times is unreliable; parse and compare them in code, as typesafe-computer-use does.
- Long, noisy states. A full page dump with hundreds of irrelevant elements lowers accuracy. Filter to the visible, actionable elements first.
- Multi-step reasoning. Jev picks a good next step, but planning a long sequence is a System 2 job; keep the plan in code or give it to a reasoning model.
- Adversarial pages. Text on a web page can try to steer the model, and TypeSafe notes that Jev does not treat state as hostile by default. Do not let an agent take irreversible actions on untrusted sites without a check.
Try it
Section titled “Try it”Both projects are open source and need a TypeSafe API key; see How to get a Jev API key. Budget-wise, TypeSafe’s launch post notes that its Doom-playing demo made ten requests a second for about $7 an hour, which is a useful upper bound for an agent that decides continuously.