02 / 06

What people are actually building with Jev

Demonstrations you can watch, open-source projects you can read, and a catalogue of every deployment reported in public in the model's first days — agents, support queues, ranking, games and robots, and a consumer slice that is almost empty.

01Demonstrations

These are other people’s builds, not ours. Each still below is a frame from the demonstration its author published, and each one links back to the original post so you can watch it run and read the thread rather than take our summary for it.

  1. Still from Gregor Zunic's demonstration: A browser agent booking a flightVideo01

    A browser agent booking a flight

    Browser Use wired Jev into their agent loop: a new action space each step, the DOM as state, and a small language model kept only for typing. The run in the clip finds flights in seconds.

    Breaking: Browser Use + Jev = Ultrafast

    @gregpr07

  2. Still from aaron's demonstration: Computer use, driven by classificationVideo02

    Computer use, driven by classification

    The screen is read with OCR and the next action is chosen from a fixed set rather than written out. Because nothing is generated, the same loop works across operating systems.

    i built computer use using @typesafeai ! it is 155x cheaper than opus 5, ~20x faster, and generalizes across OS's

    @awlevin

  3. Still from Faadil Shaik's demonstration: Super Mario Bros played from emulator stateVideo03

    Super Mario Bros played from emulator state

    Structured game state goes in and a controller input comes back, fast enough to keep up with the game. No part of this is a language task.

    got @typesafeai’s new model Jev to play Super Mario Bros.

    @faadilhshaik

  4. Still from Ziwen's demonstration: DOOM, played like a playerVideo04

    DOOM, played like a player

    TypeSafe's own demonstration, which is what pushed the model into general view. The point is latency and cost per decision, not game-playing skill.

    Jev plays doom in realtime and it looks like an actual player.

    @ziwenxu_

  5. Still from Filip's demonstration: Driving app and UI testsVideo05

    Driving app and UI tests

    Paired with opencode, judgment calls about what a screen is showing and what to tap next are made inline instead of round-tripping through a chat model.

    Testing apps with @typesafeai jev & @opencode will be soooo fast

    @Neriousy

  6. Still from Hamilton Ulmer's demonstration: A DuckDB extension that classifies rows06

    A DuckDB extension that classifies rows

    Classification as a SQL function over any CSV, Parquet file or table — the shape that makes sense when the model is cheap enough to run on every row rather than a sample.

    I made a DuckDB extension where you can use @typesafeai 's Jev to do quick classification of rows in any csv/parquet file or duckdb table

    @hamiltonulmer

A pattern runs through all of them. The set of allowed answers is built in ordinary code — legal moves, controller inputs, elements on a page, rows in a table — and the model only ever chooses between them. It is never asked to invent an option, which is why none of these systems need a validation layer behind the model.

02Open source built on it

The clearest evidence of what people are doing is what they published. Every repository below is public, and the star counts are from the day of the sweep rather than today.

Fig. 01
RepositoryWhat it doesLanguageStars
browser-use/jev-ultrafastBrowser Use's Jev-powered agent loop: a new action space each step, the DOM as state.Python1334
vinnylarouge/jevlikeTraining code for reverse-engineered Jev-like models you can train yourself.Python562
TheoLeeCJ/openjevCan we run something like Jev on a 3090 at home?Python547
jarrodwatts/jev-traderOne AI trade decision every Monad block. Jev on Kuru MON-USDC.TypeScript393
fhshaik/typesafe-marioA TypeSafe/Jev agent that plays Super Mario Bros. from structured emulator state.Python185
awlevin/typesafe-computer-useComputer use for about $0.0002 a step: OCR the screen, classify the next action with TypeSafe, click. macOS.Python149
thruwire/foremanSoftware Factory Foreman based on TypeSpace Jev modelPython100
devagrawal09/jev-reviewA staged code-review workflow and local dashboard built with TypeSafe Jev.TypeScript88
gargpratyush/jev-routerRoute to the cheapest model in claude code for your task using jev-routerJavaScript53
RomanSlack/jev-droneCamera-only autonomous drone in MuJoCo with a small judgment model (TypeSafe Jev) in the loop at 2.5HzPython43
vlad-terin/jev-browserJev-powered element selection for your agent’s existing computer-use toolsJavaScript36
jkudish/jev-mcpProof of concept MCP for Typesafe's new Jev AI modelTypeScript29
kitze/unclutterWXT browser extension: Jev-powered page clutter removal with reusable template rules.TypeScript25
devanshbatham/commit-minerClassify Git commit diffs and messages with Jev. Bug fixes, security fixes/CWEs, and change types.Rust15
ellipsis-dev/blinkCodebase search powered by Jev from @typesafe-aiTypeScript9
kieranklaassen/ruby_llm-typesafeTypeSafe structured-output provider for RubyLLM 2Ruby9
Public repositories built on or around Jev, by stars at time of capture.

03The shape of the demand

Counting the groups is more informative than reading them. Two thirds of everything reported is an agent, a queue or a pipeline. The consumer row is the one worth staring at if you are deciding what to build.

Fig. 02
How the 86 reported deployments divide between the six groups.

04The catalogue

Everything found in public in the model’s first days, grouped by the kind of decision being made. 55 of the 86 link to the exact post or comment they came from, most of them into the Hacker News launch thread. The rest were described on X by accounts whose individual posts this sweep did not capture, so they carry the handle and no link.

Rows credited to TypeSafe are the vendor describing what the model is for, not somebody else’s deployment. They are kept in because they are still the clearest statement of intent, but they are not evidence of use and should not be read as such.

Agent control and tool use (23)

The largest group by some distance. In nearly all of these Jev sits beside a chat model rather than replacing it, deciding what the bigger model should do next.

Fig. 03
#What it decidesSource
1Computer-use agent driving an operating system@awlevin
2Browser agent searching for flights end to end@gregpr07
3Reviewing each action an agent wants to take, before it acts@FaztTech
4Improving which tool a model reaches for@MichaelLee04
5Deciding when to retry a failed operation@MichaelLee04
6Context packing — whether every token really needs to be sentTypeSafe
7Proactive memory formation — did we just learn something worth keeping?iforgotmypasswo
8Memory filing — which category to store under, which class to retrieveiforgotmypasswo
9Splitting a multi-intent voice command in Home Assistantbradly
10Criteria-based model routing@eve
11Routing requests between models@DAIEvolutionHub
12Checking another model's output before it is used@DAIEvolutionHub
13Detecting jailbreaks in prompts and reasoning tracesTypeSafe
14Scoring, judging and verifying model outputsTypeSafe
15Operating a computer through the OS accessibility treepostalcoder
16Driving an AI video editor through long chains of tool callsdthedavid
17Intent detection@MichaelLee04
18Steering model behaviour mid-run@MichaelLee04
19Contradiction detection across a long conversation@MichaelLee04
20Deciding whether to send a proactive follow-up message@MichaelLee04
21Silence handling — whether a voice agent should speak into a pause@MichaelLee04
22Safety and observability tagging@MichaelLee04
23Design-by-contract enforcement around model callsfuturisold
agent control and tool use, as reported.

Customer and business operations (16)

Classification work that until now either ran on a frontier chat model at frontier prices, or on hand-written rules nobody wanted to maintain.

Fig. 04
#What it decidesSource
24Routing a support ticket to the right department@erhanmeydan
25Scoring ticket urgency on a fixed scale@erhanmeydan
26Flagging cancellation and chargeback risk@erhanmeydan
27Deciding whether a conversation should go to a human@erhanmeydan
28Reading customer anger on custom rungs: annoyed, furious, outraged@FaztTech
29Deciding which inbound emails auto-route and which need a personVercel
30Classifying an email as a cancellation request, with an urgency level@zubapita
31Predicting whether a customer is about to leave@AGTPinsights
32Deciding whether an event is worth alerting a person about@AGTPinsights
33Spam detection@AGTPinsights
34Phishing detection at population scalettul
35Fraud detection@FaztTech
36Invoice validation@FaztTech
37Billing-complaint classification18al
38Front-line support resolved without a human in the loopTypeSafe
39Factory floor routing for questions too varied for hand-written rulesTypeSafe
customer and business operations, as reported.

Ranking, matching and retrieval (13)

Where the job is to compare one thing against another and put a number on the fit. Several of these replaced an embeddings pipeline rather than a language model.

Fig. 05
#What it decidesSource
40Scoring a resume against a job description in a consumer hiring app@h_nilforoshan
41Ranking search results@GeniusGTX
42Picking the five most relevant articles out of a thousandlatteren
43Genealogy person matching on names, dates and placesvintermann
44Record linkage using household nestingRobinL
45Entity resolution, with blocking to keep the candidate set smallcamdenclark
46Asking whether a contract contains a given clauseymir_e
47Extracting key facts from energy-asset data roomsymir_e
48Zero-shot classification in place of a frontier chat modelpetesergeant
49Bulk classification of a million transcripts, previously embeddings plus cosine similarityjawns
50Classifying rows of any CSV, Parquet file or table from inside DuckDB@hamiltonulmer
51Turning very large datasets into features through map-reduceTypeSafe
52Agentic retrieval over a large corpus while the chat UI stays responsiveconsumer451
ranking, matching and retrieval, as reported.

Code and developer workflow (9)

Mostly decisions inside a pipeline that already exists — which file to open, whether a failure is real, whether a diff obeys a written convention.

Fig. 06
#What it decidesSource
53Finding the key symbols in a codebase before generation or reviewhunterbrooks
54Evaluating code reviewsnickstinemates
55Semantic linting — scoring a diff against a written conventions fileTypeSafe
56Mediating CI flakespassive
57Driving app and UI testing with opencode@Neriousy
58Decision-tree logic inside a shipped extensionnickstinemates
59RubyLLM integration@kieranklaassen
60Agents that decide, route, score and stop, via AI Gateway's evaluate()@vercel_dev
61Local validation of structured eventsHacker News
code and developer workflow, as reported.

Games, robotics and real-time control (12)

The group that best shows what a model with no text to generate can do: a decision arrives fast enough to sit inside a control loop.

Fig. 07
#What it decidesSource
62Playing DOOM in real time from structured game stateTypeSafe
63Wikiracing — choosing one outgoing link per hop to reach a target articleTypeSafe
64Game QA — instrumenting a game and emitting control inputscaspar
65Accelerating QA at a game studiostrich
66Trading bot acting on a price feed, block by block@jarrodwatts
67Robotics — limb controlFlere-Imsaho
68Robotics — general planningFlere-Imsaho
69Route findingFlere-Imsaho
70Driving rules — stop on red, go on green, stay in lanepantelisk
71Choosing the next note in a partly-written song, constrained by a grammariforgotmypasswo
72Playing Super Mario Bros. from structured emulator state@faadilhshaik
73Camera-only autonomous drone flight in simulationRomanSlack
games, robotics and real-time control, as reported.

Things a person actually sees (13)

The thinnest group, and the one to read closely if you are building for consumers. Almost everything else on this page is buried inside somebody's backend.

Fig. 08
#What it decidesSource
74A website restructuring its own layout for the visitor in front of it@furoku
75Inferring visitor context from referrer, device and history@furoku
76Deciding whether to show a promotion to a given visitor@furoku
77Re-ordering page sections by what that visitor is likely to want@furoku
78Content firewall — hiding feed posts by a natural-language rule@marcelpociot
79Running 150 synthetic personas through a product survey@ytiskw
80A personal delivery tracker reading incoming emailxixixao
81Smart-home command interpretationVladVladikoff
82Hardware that anticipates what its user wants nextTypeSafe
83EdTech product decisions, held up by EU data residencyfreddex
84A structured judgment data lake across ten human skillssaldubai
85On-device reproductions of the idea, openjev among them@harshagundal
86Stripping clutter out of a web page in the browserkitze
things a person actually sees, as reported.

05What is not on this list

No writing, no summarising, no translation, no chat, and no image or audio work. That is not an oversight in the sweep — the model produces no text, so none of those jobs are available to it. If your use case is missing from this page, the first question to ask is whether it is a judgment at all.

The second absence is time. Every deployment here is days old. Nobody has yet reported what happens after a year of drift, a model version bump, or a change in the shape of their own input data.