Jev in Practice: Browser Use and Agent Workflows

After hand-building a Jev client across six files in Spring Boot last week, the author packaged it into a Spring Boot 4 starter and recorded a video explaining how the auto-configuration works.

A roundup of three free ways to use Jev right now: the official Playground, Vercel AI Gateway selecting typesafe-ai/jev, and OpenCode Zen selecting jev-1.13-free, with a note that access is time-limited.

The plugin does not rewrite memory; it uses a small fast model to score each tool call and its output, cutting low-scoring ones and keeping the rest word for word, plus five things to check before installing.

The author tried letting Jev rather than an LLM drive the agent decision loop and got about 20% on a spreadsheet financial modeling benchmark versus ~80% for LLM agents, concluding Jev suits bounded tasks and use as a tool inside the classic loop.

A walkthrough of handing market state to Jev to judge regime, whether flow is toxic, and whether a setup is worth taking, returning a typed decision for the trading system to execute.

Live TradingView indicators are compressed into a snapshot under 400 tokens, and Jev scores typed questions in parallel to return calibrated buy, sell, or hold probabilities that drive Kelly-sized positions, with an open-source repo.

A demo shows Jev turning natural-language instructions like "open your hand, go to the ceiling, stretch as far as you can" directly into robotic arm movements, ahead of an Argentina Space event.

LangChain compared Jev with GPT-5.6 and Claude Sonnet 4.6 as agent-eval judges: Jev matched the human reviewer on every call, with up to 913x lower variance and 0.44s per call, for a full-run cost of $0.34.

The author ran openJev Verdict on a 151M base model without retraining and, after fixing three deterministic inference bugs, reached 74.9 on JevBench, explaining why Jev is replacing chat models for routing.

Locus Pro replaced tool routing with Jev, letting agents semantically search across its 3,600+ tools to decide which one to use and when.

FundaAI argues Jev suits low-cost labeling, scoring and yes/no decisions, but the evidence supports only a narrow claim and does not change the AI infrastructure investment thesis.

The author spent a Sunday building a 50-person simulated city on Jev, where residents fall in love, cheat, gossip, get sick and die, running 14 hours, 74 simulated days and nearly 11,000 requests for $1.05 across two models, and uses it to explain what Jev actually is.