What Can You Build with TypeSafe's Jev? 18 Ideas from Its First Week

Jev is a new AI model from TypeSafe AI. At first it sounds like a step backwards.

It can’t write a sentence. It can’t chat, summarize or write code.

You send it some data and a question. It answers yes or no with a probability, picks one option from a list you wrote, or gives a score on a scale you wrote. That is all it does.

So what would you build with it?

I had the same question. Then Moritz Kremb posted a thread of ten Jev projects on September 18. When I checked two days later it had about 6,800 bookmarks and 4,000 likes. A lot of people were asking the same thing and saved the answer.

This page goes through those ten and adds more I found on GitHub. For each one I explain the idea, because the idea is the part you can reuse.

Here is the short version. Builders report answers in 100 to 400 milliseconds, and TypeSafe charges $0.042 per million input tokens. So you can ask a question at moments where a chat model was always too slow: on every keystroke, every spoken word, every row in a table, every command an agent runs.

And in almost every project, code does the looking and builds a short list. Jev picks from the list.

One note before the list. I haven’t run Jev. Access is by waitlist. Every number below is the builder’s own, from their post or README.

What is Jev, and what does it answer?

TypeSafe announced Jev on September 15, 2026. It calls it a System One model, after the fast kind of thinking in Daniel Kahneman’s Thinking, Fast and Slow.

A request has two parts. The state is your data, as text or JSON. The questions are typed, and there are three types.

• Noul is a yes-or-no question. You get a probability between 0 and 1.
• Choice picks one option from your list. You get the pick and a probability for every option.
• Score places the state on a scale you define.

You can send many questions about the same state in one request. They are answered in parallel. Several projects below depend on that.

Can Jev drive a browser or a game?

Yes, as long as code turns each step into a list of legal moves.

(1) Code numbers the buttons, and Jev picks a number.

This is Jev Ultrafast, from the Browser Use team. Gregor Zunic posted it with a flight search that took 7 seconds and cost $0.0039.

On every step, code reads the page and writes a numbered table of elements: [3] combobox Where to?. Jev answers two questions in one request. Which operation, and which element.

Jev can’t type “London”. So a small language model writes text, and only when the operation is TYPE_TEXT.

(2) It can act before you finish the sentence.

Moritz Kremb’s voice browser sends every partial transcript to Jev as you speak. Each request carries about a dozen questions.

What is the intent? Which element? Is the command complete? Is this even addressed to me? Is it destructive?

Answers come back in 250 to 350 ms at about $0.0002 each. Then code decides whether to act, wait, ask or ignore. In his post, “go back” ran before he finished saying it.

Search text is handled the same way as in (1), without a second model. Code pulls candidate phrases from the transcript, Jev picks one, and it is copied as spoken.

(3) Give it the game’s state, not the screen.

Jev reads text. It doesn’t take images yet. So jev-doom-agent pulls health, ammo, position and visible monsters out of a real Doom engine running in the browser.

Jev picks a tactic from a list. A local controller turns the tactic into key presses. A low-confidence answer is shown as a fallback on screen, never passed off as Jev’s.

TypeSafe’s own launch demo played Doom too. The community lists have the same idea for Mario, Pokémon and StarCraft.

(4) It can make one decision per block. That doesn’t make it a good trader.

Jarrod Watts’s jev-trader asks Jev “buy or sell?” every 300 ms, once per block on the Monad chain. It then places a real limit order on that side. His post links a live dashboard.

The idea to take from it is the time budget. A decision and an order both fit inside 300 ms.

The README never says the bot makes money. Its sample output shows a loss of a fraction of a cent. Without a wallet key it runs as a simulation.

(5) Three demos in the thread have a video and nothing else.

Jack Cheng drew on a canvas by voice. Justin Schroeder says he rebuilt Tesla’s self-driving in under an hour. Filip tested an app with Jev and OpenCode.

None of the three links a repository, so how they work is a guess. My guess is the shape from (1): code lists the moves, Jev picks one.

For the canvas there is an open-source project with the same idea, jev-canvas. It draws on a tldraw canvas from your voice and a pointing finger.

What can it do on every keystroke?

It can rank a list by what you probably mean. Until now that job went to fuzzy matching, because nothing smarter fit inside a keystroke.

(6) A launcher can rank by what you mean.

Nader Dabit’s keystroke oracle is a Spotlight-style launcher for macOS. Type the pdf I just downloaded and the newest PDF is the top hit. Type wifi off and it picks Turn Wi-Fi Off over Turn Wi-Fi On, which a fuzzy matcher scores the same.

Code collects the local candidates. Jev ranks them on every character, with no pause added to wait for you to stop typing. 15% of 240 is worked out in code, because arithmetic isn’t Jev’s job.

It lives in jev-experiments beside about twenty other small demos. The README says they were built by Devin, the coding agent.

(7) Shell history can be ranked the same way.

jev-shell-history is a zsh plugin. It looks at your last 100 distinct commands and asks Jev which one you are typing.

So git st completes to git status. The better trick is the second mode. Type last 5 commits, and it offers git log --oneline -5 from your history.

Can it label things nobody has labeled?

Yes. You describe the label in a sentence. There is nothing to train and nothing to index.

(8) A WHERE clause can take a sentence.

pg-jev is a PostgreSQL extension by Zachi.

SELECT * FROM people WHERE jev(people, 'could work from home');

Every row is judged yes or no, 20 rows per request. jev() is an ordinary function, so it works with joins, ORDER BY and LIMIT. Put a cheap condition first, as in WHERE age > 60 AND jev(...), and the rows it removes are never sent.

His post reports 129 rows in about a second for $0.0009, and 6 ms on the second run from cache. The README measures 2,000 rows at about 3.5 seconds and $0.012.

sqlite3-jev does the same for SQLite.

(9) You can describe a label instead of training one.

Riley Brown classified 500 emails in seconds for 3.5 cents. His post is a video with no code.

Sift, in Nader Dabit’s repository under inbox-blitz, shows how such a thing works. You type “customers threatening to cancel” and all 500 emails are judged against that sentence in about 5.5 seconds.

It keeps a keyword filter on screen next to Jev’s labels. You can see where keywords fail: sarcasm, negation and quoted replies.

(10) It can label a video nobody has labeled yet.

SponsorBlock skips YouTube sponsor reads, but only after a volunteer marks the timestamps. A video uploaded an hour ago has none.

jev-skip reads the captions of the video you are watching. It asks for a sponsor probability for every 30-second slice and paints the result on the seek bar.

The author reports catching 77% of the sponsor seconds that SponsorBlock’s crowd marked across 23 videos, at $0.0008 a video. The README also says those numbers came from recorded answers through a gateway, and should be measured again on the real API.

What can it check for a coding agent?

It can sit in a hook and judge each action. It is cheap enough to do that every time.

(11) Compaction can delete instead of summarize.

When a Claude Code session gets long, old turns are replaced by a summary. A summary can lose the exact file path or error message you need later.

tamara’s fast-jev-compaction rewrites nothing. For every old tool call it asks Jev two questions. Does it still matter that this call was made? Is its result still needed word for word?

Calls that fail both are deleted. If only the first is a yes, the call stays and its result is cut short. Your messages and the assistant’s stay as written, in order.

This was the most viewed project in the thread. Her post passed 3 million views, and the repository had about 4,000 stars on September 20.

(12) Every command can get a second opinion.

jev-axi runs before each tool call. Give it curl ... | bash and it denies the command. The reason it gives: likely downloading and running unreviewed code, with a probability of 0.99.

It takes about 400 ms, which is short enough to check every command.

TechCrunch reports that an engineer at Vercel uses Jev the same way, to review commands. Answers came 5 to 18 times faster than from the model it replaced.

(13) “Done” can be checked before the agent stops.

Claude Code says “Done, tests pass.” Did anything run?

jev-belay is a Stop hook. It reads the transcript locally. Only when files changed and no check has passed since does it make one Jev call, at about $0.00002. If the “done” isn’t backed by a run, the turn doesn’t end, and Claude goes back to run the tests.

If anything in the hook fails, the turn ends as normal. The demo recording used canned probabilities, and the README says so.

What can it do for writing?

It can’t write for you. It can read, and tell you what it sees.

Thomas Ptacek’s rule for writing with an LLM is to keep none of the words a model suggests. Jev is a model with no words to suggest.

(14) You can check everything you’ve written against 21 habits at once.

Mike Taylor at Every took 27 of his own articles and added 10 written in an AI style. He asked 21 questions about each one. Does it repeat an idea without adding evidence? Does it overexplain a simple point?

That is 777 judgments. He reports under 0.7 seconds and about a quarter of a cent.

He also says it needs a proper accuracy check before anyone relies on it. I agree. A flag means a paragraph has a habit. It doesn’t tell you who wrote the paragraph.

(15) A linter can count some rules and ask about the rest.

Sniff Test checks Markdown against 15 rules. Long dashes and banned words are regular expressions. They run on your machine and cost nothing.

Other rules need reading. Does the last sentence only restate the paragraph? Is a claim hedged three times? Those go to Jev, one paragraph at a time, and only after you say yes.

It won’t rewrite anything. In the README’s words, “A flag is a sentence for you to fix.”

That is the difference from Humanizer, which edits the draft for you. How to Write with AI Without Losing Your Voice has a short list of the habits such tools look for.

(16) A citation can get a confidence score, and you still decide.

Paper Trellis’s citation verifier is for peer reviewers and authors. It pairs each sentence in a manuscript with the paper it cites, and fetches the paper.

Claude proposes a verdict and quotes the passage behind it. Code checks that the quote really is in the paper. Then Jev reads the sentence beside the passage and gives a probability for each of three answers: supports, contradicts, says nothing.

The last step is yours. The README says model output never overwrites a human verdict.

What else are people trying?

(17) A question about your house can become a sensor.

HA-Jev is a Home Assistant integration. You write a question about some devices or a room. The answer shows up as a sensor: a probability, a choice or a number.

Your automations then trigger on it like any other sensor. There is a daily token budget, and evaluation stops when it is used up.

(18) Code finds the candidates, and Jev judges them.

Zachi’s ad blocker is a Chrome extension. Code finds ad-shaped elements on the page. Jev answers one question for each: is this an ad?

The author calls it a fun project and says to use uBlock Origin if you want a real ad blocker. I include it because it shows the pattern behind almost everything on this page in its smallest form.

What can’t Jev do?

Knowing the limits saves you from a bad idea.

• It can’t produce text. TypeSafe’s own wording is that Jev “gives up string generation”. Projects that need text add a small language model, as in (1), or copy words from the input, as in (2).
• It reads text only. TypeSafe says images aren’t supported yet. That is why (3) sends the game’s state and not the screen.
• It is unreliable at counting, arithmetic and comparing dates, according to Flavio Copes’s notes from his own testing. Do those in code, as the launcher in (6) does.
• A Choice can have up to 255 options. A request tops out around 32,000 tokens, according to the README of (11).
• Access is early and by waitlist.
• The numbers are early too. Most come from one run on the builder’s machine. Several READMEs above say so themselves.

Where can you find more Jev projects?

The community lists are long and growing.

• yibie/awesome-jev had 168 projects in 13 categories when I checked.
• AnotiaWang/awesome-jev is released under CC0 and covers SDKs in about ten languages, MCP servers and open-weights replicas.
• typesafeai.app is an unofficial index with a page for each of 30 use cases.
• TypeSafe’s docs cover the three question types and its own patterns.

A list of 168 links is hard to read for one idea. It is easier to ask. Clone a list into a folder, open the folder with Claude Code or Codex, and ask which projects are closest to what your app does.

We make StashBase, a writing workspace where that folder and the agent chat sit side by side. Keyword search across the files works without any setup. It doesn’t run Jev.

I haven’t run Jev, which is in early access. The descriptions come from each project’s post and README, checked on September 20, 2026, and the numbers are the builders’ own.