What are people building with Jev?

Subway Surfers at superhuman speed, then fifty games at once. The whole run cost less than a cent.

One call returns 14 typed checks on a diff — secrets, SQL injection, deleted tests, blast radius — and code turns them into block, review or merge.

Scores 700 leads and their outreach messages in 40 seconds and flags lead-message mismatches, for nine cents. Concept preview, not shipped yet.

Jev makes the trading calls with a $10,000 budget.

Each decision takes about 0.7 s, faster than you can follow the screen. GPT-6 Astra played well but slowly.

A local CoreML model finds the UI elements, on-device OCR reads them, and Jev picks what to click. No pixels leave the Mac.

Jev classifies every DOM element as ad or not and removes the ads.

Uses the real weights to simulate how viral your post will be, with a global feed of everyone's tests.

Jev driving the computer, much faster than an LLM loop.

Jev picks the tool and arguments (web search, Wikipedia, weather, Todoist, Home Assistant) and answers instantly, with sources.

Eight questions per post, 4.2M tokens for $0.13: how-to posts got 150 median likes against an average of 44.

Observe the page, send the accessibility tree and actions to Jev, let Stagehand execute. This task cost $0.001 in a remote browser.

Talk to a real browser: the transcript goes to Jev, which answers in about 300 ms and clicks. $0.0002 per decision.

A two-armed robot where Jev handles the decision layer and plain code does the kinematics and physics. 500 ms responses, about half a yen per attempt.

A browser agent where a single Jev request picks both the action and the element to click. It runs a full Google Flights search in about seven seconds.

WHERE jev(people, 'could work from home'): 129 rows judged in ~1 s for $0.0009, and 6 ms from cache the second time.

Jev sorted 500 emails for 3.5 cents.

61 questions per post in ~1 s, fitted on 9,481 real posts: write, score, rewrite until it peaks. It picks the viral post 2 times in 3.

The beginner method lives in code and Jev decides each step: 94 moves instead of the optimal 22, about 250 ms per decision.

Jev tags the personal bits of each message and a redact() SQL function in Neon masks only those.

Jev decides which model fits each request best, and the request goes to that model.

A public site where every question gets a verdict instead of a reply. Built with Convex, aiming for a million questions.

A browser extension that collapses posts matching rules written in plain language, so fast you don't notice it.

Jev controls the player in real time; a walkthrough of what makes it different from an LLM.