An inbox that gets you. A browser that moves.
An agent that touches grass.
See what happens when AI learns to choose.
SANDBOX / MINECRAFTTHE GOOD STUFF / 01—06
Real projects. Delightfully different uses.
Watch one in action, then find your next experiment.
SANDBOX / MINECRAFTLet your agent touch grass.
Gather resources. Pick the next move. Watch a decision model find its feet in a blocky world.
How Minecraft worksOriginal JevFlow concept art · not gameplay.
Explore the Minecraft case study ↗One goal. A browser in motion.
Watch Jev pick the next click while a text model handles typing. Hello, flight-search autopilot.
How Browser Use worksBrowser Use · original project recording.
Explore the Browser Use case study ↗
Find the meaning in the mess.
“Who actually wants a refund?” Turn a crowded inbox into a short list worth your attention.
How Sift worksNader Dabit · project screenshot.
Explore the Sift case study ↗
A column header with superpowers.
Type “Urgency” or “Ready to buy?” and let the rows fill themselves. Your spreadsheet gets the hint.
How Judge Sheets worksNader Dabit · project screenshot.
Explore the Judge Sheets case study ↗SELECT name, message
FROM customers
WHERE jev(customers,
'wants to cancel'
);SQL, meet a little common sense.
Ask your tables for “customers about to cancel.” Filter and rank rows by what they actually mean.
How pg-jev worksJevFlow illustrative query · example data.
Explore the pg-jev case study ↗Keep the signal. Lose the baggage.
Give your coding agent a lighter backpack. Keep useful tool history and trim the stale stuff.
How Fast Compaction worksJevFlow illustrative preview · not a benchmark.
Explore the Fast Compaction case study ↗Made by curious people around the internet. Meet the model behind the projects ↗
A DIFFERENT KIND OF AI
A Jev AI agent puts quick decisions to work. Jev picks, scores, and sorts. Code takes action. Other models step in when the job needs planning or words.
Get to know JevA page, a message, a world state.
Tools execute. The workflow checks.
CURIOUS? GOOD.
The useful bits, minus the jargon.
An agent that uses TypeSafe’s Jev model to make decisions during a task: choosing an action, evaluating an item, or routing work. Tools execute the actions; another language model can handle planning and writing.
Right now, you can explore the community demos and visit their original projects. Try takes you to JevFlow’s early-access page for the experience you picked. Hosted agents are not available here yet.
Jev returns structured choices, scores, and truth probabilities from text-based state. In these projects, surrounding tools handle browser observations, game state, and execution. Other models can handle text generation.
JevFlow is independent. These projects were built by community developers, with author and source links on every card. Jev itself is made by TypeSafe.
No. A structured answer can still be wrong. Good agent workflows validate actions, check outcomes, and send uncertain cases for review. A compelling demo is not a reliability guarantee.
For current model access, documentation, and pricing, visit TypeSafe’s official website. Running a full agent may also involve browser, hosting, and other model costs. Visit TypeSafe ↗
YOUR NEXT “WAIT, IT CAN DO THAT?”
Help shape what comes next for JevFlow.