How to Use Jev With Claude Code, Codex and Cursor (Free Reference File)

Sebastian Bennis — Thu, 24 Sep 2026 11:41:11 +0000

TL;DR: Jev is a new kind of AI that doesn't write. It decides, in well under a second, for fractions of a cent. It's perfect for the hundreds of small yes/no calls inside an app. Your coding assistant probably doesn't know it well yet, so I wrote a free reference file you can hand to Claude Code, Codex or Cursor: github.com/sebastianbennis/jev-system-one-reference.

What is Jev, in plain English?

Most AI models are essay writers. You ask a question, they write an answer, and your code has to figure out what the words mean.

Jev, released by TypeSafe AI on 15 September 2026, is more like a multiple-choice machine. You give it:

Some text or data (a support message, a web page, a list of options)
Questions with fixed answer types

It answers every question at once, and each answer comes with a probability. There are only three question types:

Type	The question it answers	Example
Noul	Is this true?	"Is this message spam?" → `0.97`
Choice	Which one of these?	"Which team should handle this?" → `billing`
Score	Where on this scale?	"How frustrated is the customer?" → level 1 of 0–2

Your code reads the answer and decides what to do. That's why people call it the "if statement" for AI.

Why vibe coders should care

Speed and cost. TypeSafe lists the price at $0.042 per million input tokens, with output free.

When I tested it myself, one request with 7 questions came back in 0.62 seconds and cost about $0.00005.

Other builders have posted bigger results (these are their numbers, not mine):

Browser Use built a browser agent where Jev picks every click. It completed a Google Flights search in about 7 seconds (repo).
Matthew Berman had it break down 724 live ads in 40 seconds for about 9 cents.
Elvis (@elvissun on X) had it triage 384 news headlines for 15 brands in under 25 seconds.

The combo: a big model for thinking, Jev for deciding

Jev can't write text. It can't draft an email, write code or explain itself. That's not a flaw; it's the design.

The pattern the community has settled on:

🧠 A big model (Opus 5.5, GPT-6 Sol, whichever you use) does the thinking and writing.
⚡ Jev makes the fast, repeated decisions in between: route this, flag that, is this relevant, is this safe to send.

If your app is paying a frontier model to answer yes/no questions hundreds of times a day, that's the part Jev can replace.

Why your AI coding assistant gets Jev wrong

Jev is brand new and its API has already changed. When you ask Claude Code or Codex to "add Jev to my app", they often reach for outdated or guessed patterns. The common mistakes:

Using an old field name. Early versions used document. The current API uses state, and document now fails validation.
Using jev-latest instead of a pinned version. The alias moves when a new version ships, which can silently change your results. Pin a version such as jev-1.13.0.
Treating "confidence" as permission. A confident answer can still be wrong. Your code should decide what happens, with a human review step for anything important.
Asking it to do maths or compare dates. TypeSafe's own docs say it's unreliable at counting, arithmetic and date comparison. Do those in code.
Asking one giant question. "Is this ticket urgent and about billing?" works worse than two separate questions. Jev answers them in parallel anyway, so splitting costs almost nothing.

The fix: give your assistant a reference file

I wrote an open-source reference designed to be read by your AI assistant, not just by you. It's one Markdown file containing:

✅ The current API shape, with a complete working request and example response
✅ The three question types and when to use each
✅ Known failure modes and how to avoid them
✅ How to combine Jev with a big model safely
✅ Copy-paste prompts for starting a project and for continuing it in a new session

How to use it in 3 steps

Download Jev_System_One_Reference.md from the repo into your project folder.
Tell your assistant to read it first. For example:

Read Jev_System_One_Reference.md, starting with Section 0.
Then build a small support-ticket router in Python that uses Jev
to pick the right team and flag urgent messages.
Include a mock mode that works without an API key.
Never put the real key in the code; read it from the environment.

Review what it built. The README includes a checklist of what to check before calling it done.

That's it. The repo's README has a full worked example and a prompt for continuing work in a fresh session.

I asked Jev to pick this article's title

I didn't pick this title myself. I wrote five candidate titles and seven opening lines for the X post, then sent them all to Jev in one request with 16 questions. It answered in 0.6 seconds for about $0.0001.

It picked this title with 72% of the probability. For fun, I also included a fake hype line claiming "Opus 5.5 and GPT-6 Sol combined with Jev changed everything for my app". I haven't built that, and the facts I gave Jev said so. Jev ranked it dead last and flagged it as unsupported.

That's the point: Jev judges exactly what you ask it to, against exactly what you give it.

FAQ

Is Jev free?
No, but it's very cheap: $0.042 per million input tokens with free output, at the listed price. Some gateways such as Vercel and OpenRouter also offer access.

Does Jev replace ChatGPT, Claude or GPT?
No. It can't write. Use it alongside a big model for the quick decisions.

Can Jev hallucinate?
It can't return an answer outside the options you define. But it can still pick the wrong option confidently, so test it on your own examples.

Does this work with Cursor, Codex and other assistants?
Yes. The reference is a plain Markdown file, so any assistant that can read files in your project can use it.

Is the reference official?
No. It's an independent, MIT-licensed community reference, not affiliated with TypeSafe, Anthropic or OpenAI. It links to TypeSafe's official docs as its sources.

Prefer the short version? Here's the thread on X.

⭐ If this saves you a debugging session, a star on the repo helps other vibe coders find it. What would you build with a 0.6-second AI decision? Tell me in the comments.

DEV Community: Sebastian Bennis