Jev in Practice: Browser Use and Agent Workflows

Rob Shocks demonstrates in the playground how Jev returns typed decisions and calibrated probabilities, and gives speed and cost figures of about 100ms per call and $42 per billion tokens.

The author open-sourced a self-hosted HTTP API wrapper that turns TypeSafe Jev's decision primitives into simple GET requests, so users can try Jev by pasting a URL into the browser address bar.

The author contrasts chat agents with Jev, which takes state and returns a typed choice, score or bool in about 70-500ms with no output token charge, and says LangChain is already trying it for model routing while Vercel reports about 13% of paying Gateway teams tried it on day one.

The author used Jev to run automated playtests of Nightshift Hollows and Starfall Salvo, listing which testing mistakes mislead balance judgments and sharing a reusable starter workflow.

The ex-OpenAI founder of Jev says Claude Code still belongs to the assistance era and that Jev is the next layer, 200x faster, 400x cheaper with zero hallucination, tracing a path from prompts to agents to loops to graphs to Jev.

The author had an agent research Jev and concluded that many people err by using an LLM directly as a classifier, when the right use is to have the LLM generate a classifier.

For people who keep asking what Jev is: it is not an LLM but a type-safe decision model that picks an answer from options you provide, using choice, score and bool tools, suited to routing and filtering.

The author lists five concrete uses: model routing, tool-call risk checks, inbox and support triage, stopping agent loops, and handing work to a human, noting that permissions are still enforced by code.

A developer built a mood tracker where Jev scores each sentence against six feelings, and the strongest feeling colors a matching section of a flower as you type.

A copyable prompt has a coding agent audit the whole workspace and, per project, find tasks Jev could take over such as organizing information, reviewing content or checking another agent's work, and define the inputs, references, question and output format.

The author wrote a short post on why he is excited to experiment with Jev, with the core idea that classifier-style decisions are enough.

Buying and selling MON/USDC every block, Jev made 1,838 decisions in under 10 minutes at 85–88% confidence yet lost 5.2%; adding a rule to flatten below 0.55 confidence turned the run into +4.3%.