Jev in Practice: Browser Use and Agent Workflows

A new privacy-focused iOS search browser that uses Typesafe's System One models plus result aggregation to produce answers, and can also be used for agent and browser-use cases.

Spring AI now integrates with Jev: send it a state and typed questions and get numbers back with a confidence in about 300 ms, usable for judging, self-refinement, guardrails, RAG triage and tool selection.

The author scanned GitHub and picked 40 open-source projects that actually put Jev into browser agents, coding agents, MCP, databases, model routing and security systems, arguing that using a large model to decide left or right wastes tokens.

Laya is an open-source decision model that runs locally and supports yes/no, scoring and multiple choice; in a 30-second Snake benchmark against cloud-based Jev it reached 86.5 versus 3.2 decisions per second, but its context is only 512-1024 tokens.

A breakdown article explains how Jev engineering turns an agent stack into a real control system, making decisions up to 193x faster and 444x cheaper.

The author ranks the options by preference, puts the newest one, Jev, first, and is unsure whether to use it as their coding agent this week.

The author breaks down browser-use's jev-ultrafast example, where Jev predicts the next browser action from the previous action, task goal and structured DOM state, handing text input to an LLM, completing a booking task in 7.1 seconds.

The author open-sourced jevmail: strip out the demo animation and it is a self-hosted email client running on Vercel AI Gateway, Jev and the Gmail API, with the GitHub repo included.

In the demo, each market must pass three Jev checks (worth a look, how mispriced, is the bet too big), and only 4 of 100 markets reached the expensive large model, at about 42 cents per 10,000 decisions.

A demo shows scattered files being grouped into a folder by meaning as you type its name, using the TypeSafe Jev API for file matching.

JevHarness lets an LLM generate a task-specific decision flow once and freeze it, raising win rate on a Pokémon task from 25% to 75% over five iterations.

Moss AI describes Jev as doing one thing: given a situation and a set of options, it returns a decision with a confidence score in milliseconds, which it says matters for people building trading agents on the Moss marketplace.