Jev in Practice: Browser Use and Agent Workflows

Six of the listed repos center on Jev: official skills, the foreman software factory on Jev, the jev-sift batch classification plugin, a latency-focused demo, tool-call compression and browser cleanup.

The author fed the shadcn component library and custom style rules to Jev, which composes fully responsive interfaces in 1-2 seconds at almost zero cost.

Jev is used for inbox sorting, lead scoring, content job assignment and three pre-publish checks, with high-confidence calls passed automatically and only low-confidence ones reviewed by hand.

A developer is letting Jev make all trading decisions in an Alpaca paper account, an experiment that is still in progress.

A weekly crypto AI report notes that Typesafe's Jev is now available for free in beta on the Venice API, alongside news on private AI and robotics.

In an open-source Minesweeper experiment, Jev completed decisions in 8 seconds for about $0.003, far faster and cheaper than Fable, Astra and Grok.

The author lists six ways Jev makes a cheap judgment first and only then decides whether to call a large model, covering code retrieval, email triage, tool calls, and RAG filtering.

The former OpenAI researcher explains why Jev does not generate text and only outputs judgments, and why RLHF is becoming obsolete.

A demo of using Jev as a decision front layer: it judges which emails are junk and which need action, screening 1,700 instantly, and the same approach applies to leads and tickets.

An introduction to jev-1.13.0's Choice, Score, and Noul question types, its $0.042 per million input tokens with free output, and two-minute steps to wire it into an agent loop.

The author explains what Jev is and how people are experimenting with it, then builds a mini version tailored to making decisions in his Sonic-inspired side-scroller.

TypeSafe's Jev, a model from former OpenAI researcher Diogo Almeida that only makes decisions and never generates text, is shown picking skills and subagent models inside Claude Code with calls of 70–500ms.