Jev in Practice: Browser Use and Agent Workflows

Jev is now available on ZenMux, emphasizing typed outputs with calibrated confidence for classification, scoring and routing rather than chat text.

The author says someone is using Jev to build a trading bot and is training a brain to play games with Jev, with an on-chain bio-brain planned on arc.

An open-source Chrome extension uses Jev to classify each reply into promotion, engagement bait, off-topic or AI filler, collapsing matches into a gray bar with category and probability and allowing one-click misjudgment feedback.

The author counters claims that Jev cannot generate text by feeding it a list of candidate next tokens in an autoregressive loop, which he says kind of works.

An open-source browser automation tool that uses Jev to make structured choices over DOM elements, taking about 4 seconds and costing about $0.0016 per Wikipedia navigation, and can run as an MCP server.

A shared tip on a Jev clipboard workflow: copy all your personal details first, then let Jev decide which field to use when filling out a form.

In x402 payment flows, an agent discovers candidate tools via Bazaar and a Jev instance scores whether each tool fits the task and whether its price is worth it, with the threshold controlled by code.

Called as typesafe-ai/jev, it takes shared state as input and returns typed Choice, Score, or Boolean answers with probabilities and no string parsing, suited to agent routing and guardrails.

After reading the paper, the author argues the fine-tuned model was trained only on sales-call conversion rates, is about 50% less accurate than Jev, and that Jev generalizes to any input.

Someone built a clipboard tool that stores all personal info up front and lets Jev decide which field each detail goes into when filling out forms; the author notes an LLM plus browser use could do it too, but finds the Jev use clever.

Jev was combined with Playwright-controlled Chrome into jev-browser, a general browser automation skill that runs across agents, with demos on Antigravity CLI and Codex for finding related articles and job openings.

The author passes everything written so far as state, asks Jev for the single most likely next character, appends it, and repeats, which is slow and costs more with every character.