Jev in Practice: Browser Use and Agent Workflows

Ian Nuttall shows how installing the typesafe-ai skill lets an agent find slow, expensive LLM calls in a project and suggest Jev replacements.

The author built an experimental row-level security policy linter for Supabase with Jev, with a live demo and walkthrough video.

A Jev-driven document router checks each PDF page, sends only the pages that need OCR to it, and extracts the rest locally to cut cost and speed things up.

At $0.042 per million input tokens with free output, Jev can classify second-brain notes, run content workflows, analyze posts and review SEO articles.

The official typesafe-ai skill covers architecture, routing, extraction, reranking, verification and API integration, while building-with-jev-skill focuses on Choice/Score, confidence thresholds and debugging.

Wasp co-founder Matija's 45-second demo shows Jev turning code review into a millisecond decision call that returns typed checks and probabilities, cutting the cost of 1,000 PRs from about $14.5 with Opus 5 to 7 cents.

The author notes Jev's pricing structure, where you pay for the question but the answer is not billed, and says the number is worth turning over.

Inspired by Jev, sarvam-jev reads typed decisions directly from sarvam-1's logits instead of generating JSON token by token, running 8.8x faster with the same weights and output that cannot be malformed, with Hindi, Tamil and Bengali demos in the browser.

A long thread explains Jev's positioning and community uses: give it a state plus closed questions and it returns Choice, Score, or calibrated yes/no in 70–500ms, with examples like voice-controlled browsing, a Monad trading bot, 724 ad analyses, and 3,282 tweet reviews.

The author shares two approaches: using Jev as an LLM tool that orchestrates MCP calls, or making Jev the outer loop with an LLM as a tool, relying on MCP output schemas so Jev can build a workflow AST.

The author plans to use Jev as a fast browser lane, let local models handle text, and escalate by confidence so local agents run faster and cheaper.

Former Apple engineer Tamara open-sourced fast-jev-compaction, which uses the small Jev model to score tool calls and delete useless records; Theo argues deleting history breaks reasoning and invalidates prefix caching, and the author's own test found that deleted error logs make the agent repeat the same mistakes.