Meet Jev
What is Jev? Meet the AI behind four browser games
Jev is an AI model from TypeSafe AI. CrazyJev is an independent website where you can see it take part in browser games. That distinction answers a common first question: the model and the place where you play are different things.
On the site, you can watch Jev look for furniture, choose a future section of a running course, respond to a signal, or select objects from a description. Each game gives a small decision a visible consequence. You do not need to connect an API before trying one.
What does Jev actually do?
TypeSafe AI describes Jev as a System One model built for decisions inside software. An application supplies a question and the shape of an allowed answer. The model returns a typed decision with probability or confidence information that software can use.
Think about a runner approaching the edge of a roof. The game needs to decide what sort of section comes next. It can ask Jev to choose from the permitted route types. The game then builds the platforms using its own rules. The choice comes from the model; the movement and collisions still belong to the game.
Here is how the names fit together:
| Name | Role |
|---|---|
| TypeSafe AI | Develops the Jev model |
| Jev | Makes a decision from the supplied input |
| CrazyJev | Provides games in which you can observe those decisions |
For model documentation and developer access, use the provider's official resources. For a playable introduction, the games offer a more direct starting point. They show specific uses rather than every possible capability of the model.
Where can you see a decision happen?
In Room Rush, you choose a room and describe what you want. A small bedroom needing bedside storage is a useful example. Jev participates in browsing and product selection. The computer inside the room shows the browsing session. A timer, shopping list and action log help you follow it.

This screenshot shows the starting screen. Once the round begins, watch the sequence: a search, a product page, then any confirmed find. A promising search result alone does not establish that an item reached the final list.
Skyline Run uses a different division of work. You control movement, jumping and dashing. In live mode, Jev chooses later route types. The local game creates the terrain. You remain responsible for getting across it. The model does not press the jump button for you.
Thing Finder makes the decision easier to isolate. Describe “things I would wear in winter.” Jev selects from a fixed collection of 104 emoji objects. Add “without shoes” and the game asks for a new selection. It does not search the web or invent extra objects. Dragging and collisions run in your browser.
What should you look for in a round?
Begin with a request you can judge. “Find storage for a small bedroom” gives you something to compare with the result. A vague request such as “make it nice” leaves much more open to interpretation. Neither wording guarantees a good result, but the specific one makes the outcome easier to discuss.
Change one condition at a time. Keep the room the same and ask for a different purpose. In Thing Finder, keep the original description and add one exclusion. You can then see which objects changed, instead of trying to explain several changes at once.
Separate the decision from its execution. A model may return a choice while a browser action still needs to finish. A store page may load slowly. Image generation may be unavailable. An empty final list means the round did not retain products; it is not a completed shopping plan.
Treat timing with similar care. A visible request time can include network and service processing. One round cannot tell you how fast the model will be on every device or connection. The reaction game also adjusts its score according to a disclosed game rule.
A useful first session ends with one clear observation, even if the result is imperfect. Note the request and what actually appeared.
Questions people ask first
Is Jev a chatbot?
The Jev discussed here is aimed at software decisions. CrazyJev introduces that idea through actions and choices in games. Its main experience is not a long conversation window. If you want to understand the underlying model beyond these examples, check the provider's documentation.
Does every interaction call Jev?
No. Skyline Run has a local practice mode. Moving emoji objects in Thing Finder also happens locally. Choose live Jev mode when you want to observe model participation, and read the decision panel. A moving screen alone does not prove an AI request took place.
Is CrazyJev an official TypeSafe AI product?
CrazyJev is an independent game website and does not represent TypeSafe AI. It provides interactive experiences around Jev. For a practical next step, read the Jev game guide, pick one activity, and watch where the model's choice changes what you see.