Community demos
Jev demos: what people built with Jev, with the videos
In the week Jev launched, developers posted short videos of what they had built with it: games it plays, browsers it drives, lists it ranks. These are 26 of those Jev demos, grouped by what Jev does in each, with the author's own video and a version you can run here.
Quick answer · verified 2026-10-11
What do the Jev demos have in common?
Every one of them asks Jev for the same small decision over and over, more often than a chat model could keep up with: a move, a click, a rating, a yes or no. Most pick an option. None asks Jev to write anything; where text is needed, an LLM sits beside it.
| Demos | 26 — posted 17–20 September 2026 |
|---|---|
| Pick an option | 17 demos — Choice |
| Rank or rate | 5 demos — Score |
| Yes or no | 4 demos — Noul |
The pattern on each card is our reading of the post, not the author's code, and the link at the bottom of a card goes to the same pattern on this site. Videos load from X as you scroll to them.
Jev game demos: Jev picks the move
A move is a Choice among legal options, and a game asks hundreds of them.
Mario that forks its own timeline
Choice@theappcypher · 2026-09-18
Jev chooses every move. When Mario dies, the sandbox VM forks into four timelines and the one that survives carries on.
Game benchmark →Doom with branching futures
Choice@toksdotdev · 2026-09-18
Jev picks each move while a harness runs several attempts from one checkpoint; an LLM reviews them and play resumes from the best.
Game benchmark →Codenames against Jev
Choice@uwwgo · 2026-09-20
A person plays Codenames with Jev. The author's point is the price: cheap enough to put inside a consumer game.
Wikirace →Several Jevs playing Catan
Choice@sachpatro97 · 2026-09-18
Every player is a Jev. After a while they stop trading and pass their turns: worth knowing before you build negotiation on a decision model.
Connect Four →Keeping a city alive on a whale
Choice@gokayfem · 2026-09-18
GPT-6 Astra wrote the world, Jev chose each round's action, and a generative model rendered one decision per round: Jev as the decider inside a bigger pipeline.
Chess →
Jev agent demos: browsers, desktops and a robot
One Choice per step over the actions on offer, with an LLM kept only for typing.
Browser Use with Jev choosing each action
Choice@gregpr07 · 2026-09-17
Each step offers a fresh set of page actions and Jev picks one; a small LLM only types. Open source, and in our showcase.
Author reports: flight search in 7 s, $0.0039 in total
Tool selection →Notte cloud browser sessions
Choice@nottecore · 2026-09-18
Type a task and Jev picks each action from the page's action space in a cloud browser, with a step-by-step replay afterwards.
Tool selection →Computer use at decision speed
Choice@savboj · 2026-09-17
A desktop agent where Jev makes the clicks. The author calls it a hundred times faster than an LLM; we have not measured that.
Jev in agents →Fresh app flows, captured on request
Choice@omarjpeg · 2026-09-18
A browser agent that collects current screens of any app flow you ask for, workable now because each step is fast.
Jev in agents →A litter-picking robot, in simulation
Choice@metrox_eth · 2026-09-18
Jev picks the target from the instruction, cans and then bottles, and the robot arm does the rest. Real Jev decisions, replayed in simulation.
Tool selection →
Ranking and scoring demos
A Score per item turns a list into an order.
Spreadsheet columns that rate rows
Score@dabit3 · 2026-09-18
Type a header such as Urgency and every row is rated from no follow-up to urgent while you type.
Author reports: about 100 ms a row
Ticket triage →3,000 kids' snacks, scored
Score@nikunj · 2026-09-18
Every snack in a catalogue is scored on several criteria at once.
Author reports: 3,000 items in 28 s for $0.11
Product tagging →Hacker News, re-ranked by six sliders
Score@VisheshBaghell · 2026-09-17
Instead of one front page for everyone, six sliders, technical depth among them, re-rank the stories for one reader.
News filter →Context compaction without a summary
Score@tamarajtran · 2026-09-17
Rather than summarising an agent's history, Jev scores every tool call for relevance and drops the rest, so compaction is instant.
Jev harness →Searching an inbox by intent
Score@dabit3 · 2026-09-18
Gmail search where each message is judged against what you meant, with embeddings pulling candidates first in a large inbox.
Reranking →Construction plan sets, classified
Choice@hari_trinay · 2026-09-19
Every sheet of a civil or building plan set gets a type; the author reports the same answers as the GPT-4.1 pipeline it was tested against.
Author reports: 26 sheets in 2.9 s for $0.0052
Invoice classification →
Filter and gate demos
One yes-or-no question per item, fast enough for every element on a page.
An ad blocker that reads the page
Noul@iam_zachi · 2026-09-17
A browser extension that asks of every page element whether it is an ad, and removes the ones that are.
Content moderation →Android notifications, sorted
Noul@twitte_ethan · 2026-09-18
Listens to app notifications and text messages and flags the marketing ones, without deleting anything.
Phishing screening →An audience of 100 personalities
Noul@legitamit · 2026-09-17
Each of a hundred characters makes its own Jev call every time you speak, to decide whether it is bored of you yet.
Playground →Prompt difficulty, judged before you send
Noul@k_grajeda · 2026-09-18
When you stop typing, Jev judges whether the prompt is simple, and the interface offers a fast mode if it is.
Model routing →
Interface and routing demos
What to show, what to ask next, which model should do the work.
Live outfit changes from what you say
Choice@nailthy62 · 2026-09-19
Jev reads the transcript and what you are wearing and picks the next item from your closet, for a virtual try-on.
Author reports: $0.0011 and about 620 ms per decision
Tool selection →Model routing for image and video
Choice@higgsfield_ai · 2026-09-18
Jev reads the prompt and picks the generation model that fits it best on Higgsfield.
Model routing →A form that decides its next question
Choice@tamirspiritt · 2026-09-18
JevForm branches on the answers so far and chooses what to ask next, in place of hand-written if/then logic.
Tool selection →A 3D character's face, ten decisions a message
Choice@john_bortotti · 2026-09-18
Mouth, brows, eyes, gaze and body are each a decision, composed live into one reaction instead of preset expressions.
Playground →Generative UI in milliseconds
Choice@ctatedev · 2026-09-18
json-render with Jev choosing components and actions from your own design system, so the interface appears as you ask.
Jev in agents →ai-cli: Jev from the command line
Choice@ctatedev · 2026-09-17
One global npm install gives any agent harness yes-or-no questions, choices and scores against your criteria.
Jev MCP →
The numbers in the Jev demos, against our measurements
Five authors put figures in their posts. They time a whole product, not one call; beside our measurements they still show where the time and money go.
| Demo | Author reports | What that means |
|---|---|---|
| 3,000 kids' snacks | 28 s, $0.11 | About $0.00004 an item, in line with the $0.000015 to $0.00003 per call we measured. |
| Construction plan sets | 26 sheets, 2.9 s, $0.0052 | $0.0002 a sheet: sheets are long, and Jev bills input tokens only. |
| Spreadsheet ratings | about 100 ms a row | Close to Jev's own server time of 70 to 81 ms; our median call from this site's servers is 165 ms with the network. |
| Browser Use flight search | 7 s, $0.0039 in total | A whole task, including the LLM that types; the decisions are the cheap part. |
| Live outfit changes | $0.0011, about 620 ms a decision | 35 to 70 times the per-call prices we measured, so most of it is likely the speech and vision steps around the decision. |
Our figures come from the twelve measured use cases, the server timings in jev-measured and this site's call log. The lesson for your own build: budget for Jev by the length of what you send, and expect the rest of the product to dominate the latency.
Build a Jev demo of your own
Most pattern pages linked from the cards have a Run this example box that needs no key. For your own idea, write the question in the playground, then take it to the Jev API with a free key. If you have built something, open an issue on jev-measured with the post or the repository, and say which decision Jev makes in it; open-source projects also go in the Jev showcase.
Jev demos: common questions
Are these official TypeSafe demos?
No. Each one was built and posted by its author, mostly in the week Jev launched, and is embedded here from X with X's own player. Jagent is independent of TypeSafe AI and of the authors; the descriptions and the patterns are our reading of each post.
Where do the videos come from?
From X: each one is the author's own post in X's player, loaded as its card scrolls into view, and X sets its own cookies when it loads. Nothing loads from X before that, and if your browser blocks X the card links to the post instead.