Jev in Practice: Browser Use and Agent Workflows

The author shows how Jev decides which model and effort level to use based on the user's prompt, embedding that decision layer into external harnesses like Claude Code.

A developer is building a demo with Typesafe Jev, arguing its main value is speed and intelligence for situations that today need a person to decide on the spot, such as robotics and live entertainment.

A roundup of the most-watched Jev repositories on GitHub, where the common pattern is using Jev as a fast decision layer around existing systems rather than as another LLM.

The author calls Jev's /api/alpha/decisions endpoint via OpenRouter and measures latency, cost and Japanese accuracy, and discusses using it to route messages between multiple agents.

Covers Jev's choice/score/boolean decision primitives, the free Vercel quota and setup commands, and lists projects already built on it such as browser-use, computer-use and code review tools.

The author built a trolley problem game where three models choose in scenarios like self-sacrifice and sacrificing a relative; GPT-6 and Jev jumped, while Claude refused.

This Claude Code mod uses Jev to classify the subagent model, main model and effort level for each request, and installs with a single npx command.

The developer uses Jev as a classifier to scan source code against vulnerability categories and the Solodit checklist, taking about 7 seconds and 7 cents for a 12-file repo as a pre-audit check.

Jev only makes decisions rather than generating text, with Browser Use and email classification tests: a 7.1-second flight search and 1,700 emails classified for 18 cents.

A locally run MCP server that uses TypeSafe Jev to give coding agents like Claude Code, Codex and Cursor structured code quality scores.

Jev is now live on Halofy, and the author used Claude plus Jev with a simple prompt to build a flight search plugin the whole team can reuse.

The author argues Jev is a decision model suited to tagging, classification, routing, browser and computer use, and tool selection, and that LLM plus Jev works better than replacing the LLM.