What Can Jev Do? Use Cases, Demos and Real Projects
Last checked · Independent guide, not affiliated with TypeSafe AI
Jev makes fast judgments about text that your code acts on: classifying and routing messages, moderating content, guarding LLM prompts and outputs, ranking or filtering search results, and choosing the next action in agents, games and browser automation. It cannot write text, do reliable math or look at images.
The easiest way to decide whether Jev fits a job: could a competent person make this call in a second or two by reading the input, and would the answer be one of a few options, a yes/no, or a rating? If so, Jev can probably do it for a fraction of a cent. If the job needs writing, calculation or a long chain of reasoning, it cannot.
The main use cases
Section titled “The main use cases”| Job | Question type | Example |
|---|---|---|
| Classify and route | Choice | Which team gets this ticket; which handler or model a request needs |
| Moderate and filter | Noul | Is this spam, abusive, off-topic, or a prompt-injection attempt? |
| Guard an LLM | Noul, Score | Is this prompt a jailbreak attempt; how harmful would complying be? |
| Rate and prioritize | Score | Urgency, severity, customer frustration, lead quality |
| Rank and re-rank | Choice, Noul per item | Pick the most relevant passages for retrieval; drop irrelevant ones |
| Extract from a closed set | Choice | Which of these candidate dates or amounts is the one being asked about |
| Choose the next action | Choice | Which button an agent clicks; which move a game bot makes |
| Verify | Noul, Choice | Does the source document support this citation? |
TypeSafe’s documentation describes four reusable patterns behind most of these: asking many questions at once (speculative fan-out), gating actions on confidence, building a composite score from several small scores, and routing by intent. Its cookbooks include re-ranking legal passages, checking citations, screening LLM traffic, classifying SEC filings into 75 industry groups and picking one skill out of 182 for an agent.
What people built in the first week
Section titled “What people built in the first week”- Browser agents. Browser Use released Jev Ultrafast, which picks each click and form field with Jev and reported a Google Flights search in 7.1 seconds.
- Computer use. An open-source Mac agent reported about $0.0002 per step by using OCR plus a Jev Choice instead of screenshots to a large model.
- Model routing. A Show HN project used Jev to decide which coding model should handle each task. TechCrunch quotes Armin Ronacher, CTO of Earendil, suggesting routing as a natural use, since an LLM would be too slow and expensive for the job.
- Safety classification. A Vercel engineer told TechCrunch that swapping an OpenAI model for Jev in a command-safety classifier made it five to 18 times faster.
- Code review. A Show HN project built a read-only command-line tool for semantic code review on top of Jev.
- Games. TypeSafe’s own demos include a bot playing Doom at ten decisions a second (about $7 an hour) and a Wikipedia-racing agent choosing among hundreds of links per page. Community projects followed with chess and Mario.
Real-time and high-volume work
Section titled “Real-time and high-volume work”Jev’s speed changes what is practical. At a few hundred milliseconds per call, you can put a model decision inside a request handler, a game loop or a UI interaction. At $0.042 per million input tokens, you can run it over every message, log line or document you have. TypeSafe describes this as map-reducing over large datasets: turning text into features that ordinary code or a classical model can use.
What Jev cannot do
Section titled “What Jev cannot do”- Write anything. No replies, summaries, translations or code. Pair it with an LLM when text is needed. See Can Jev generate text?
- Math, counting and dates. TypeSafe’s own list of known weaknesses says to keep arithmetic and date comparison in code.
- Images, audio or video. Input is text only. Convert other media to text first.
- Long reasoning chains. Questions with several hops of indirection are answered less reliably. Break them into simple questions and combine the answers in code.
The full list, with workarounds: Jev limitations.
Can Jev control robots or drive a car?
Section titled “Can Jev control robots or drive a car?”This came up repeatedly after the Doom demo. Jev can choose among actions described as text, quickly, which is part of what a controller does. But it reads text, not camera frames, it has no notion of physical safety guarantees, and nothing TypeSafe has published tests it on physical systems. Treat robotics and driving as research territory, not a supported use.
How to start
Section titled “How to start”Pick one decision in your code that is currently a brittle rule or an expensive LLM call, write it as a Noul, Choice or Score, and compare results on a few dozen real examples. Your first Jev call shows the request, and Jev pricing shows what it will cost.