How to Use Jev AI: Four Ways to Try It and Ship It, Step by Step

You open Jev expecting a chat box. You type "hello", or paste a paragraph and ask for a summary, and nothing useful comes back — or you never get that far, because the first three guides you found still tell you to join a waitlist.

Both problems have the same fix. Jev is not used the way you use ChatGPT, and as of this week you do not have to wait to use it. (If a plain search for "jev" sends you to medical pages, that is JEV, the Japanese encephalitis virus. This guide is about the AI model.)

We run jev-ai.org, a playground, batch workbench and API built on the Jev model, and the steps below are ordered around the mistakes that most often break a first call. This guide is built from TypeSafe AI's own quick start and model docs, checked on September 30, 2026, plus the requests we measured on our own account. By the end you will have run a real decision, know which of the four ways in suits you, and have the exact request to paste into your code.

Last updated September 30, 2026. Jev is still in early access; access rules, limits and prices have changed several times since launch, and the official pages linked below are the source of truth.

If you are still working out what Jev AI is and whether it is the right kind of model for your problem, start there. This page assumes you want to use it.

How to Use Jev AI: The Five-Step Version

1. Pick one decision your software already makes: route a ticket, grade an answer, flag a risky message.
2. Paste the text that decision is about into the state.
3. Ask typed questions about it — a yes/no probability, a pick from your labels, or a position on your scale.
4. Read the probabilities, not just the winning label, and choose a threshold.
5. Send the same request from code once the answers look right.

The rest of this guide is those five steps in detail, plus the part most tutorials leave out: what to do when the first call comes back wrong.

Before You Start: Jev Answers Questions, It Does Not Chat

Every Jev request has two parts. The state is the thing being judged — a support ticket, a product review, a JSON record, an agent's last action. The questions are what you want to know about it, and you declare each answer's shape up front. Jev never writes a sentence back. It returns one typed answer per question, with the probability behind it.

That is why a chat-style prompt fails. "Summarise this" and "what do you think?" have no declared answer shape, so there is nothing for Jev to return. The three shapes it does support, from TypeSafe's introduction to the primitives:

You can mix all three in one call. TypeSafe evaluates every question in parallel and in isolation against the same state, so a fifth question costs almost nothing extra in time.

Where Can I Try Jev AI? Four Ways In, Compared

There are four practical ways to use Jev today. They all run the same model; what differs is what you need before your first answer.

What other how-to guides get wrong right now

Most "how to use Jev" pages were written in launch week, and the access picture has moved three times since. As of September 30, 2026:

• There is no waitlist. TypeSafe dropped it on September 20, paused new signups a day later under demand, and reopened them on September 27. The full timeline is in our Jev waitlist guide.
• The $5 signup credit is gone. TypeSafe says new signups no longer get free credits and that it hopes to bring them back. Guides promising "$5 free on signup" are out of date.
• Rate limits have changed since launch. TypeSafe's models page now lists 100K tokens per second and 40 requests per second, and says the numbers are still adjusting.

If you only want to see whether Jev fits your problem, the browser playground is the shortest path. Everything below uses it, and step 5 shows the equivalent call for every route.

Step 1: Pick One Decision You Already Make

The fastest way to learn Jev is to aim it at something real. Good first candidates share three traits: a person could make the call in a few seconds, the answer is one of a known set, and you already have examples lying around.

• Which team should own this support ticket?
• Does this model answer actually follow from the cited source?
• How frustrated is this customer, on a four-step scale?
• Is this agent about to take an action it should not?

Rule of thumb: if you would need a paragraph to explain your answer, the question is too big. Split it before you start.

Step 2: Open the Playground and Paste Your State

Open the Jev AI playground. It loads with nine ready-made scenarios — ticket triage, content moderation, intent routing, review scoring, lead scoring, an LLM guardrail, a citation check, resume screening and a blank custom one. Pick the closest, then replace its text with a real example of your own.

A few things worth knowing on this page:

• Editing needs no account. You can change the state, the questions and the model without signing in.
• Running does. Press run and you will be asked to sign in; new accounts get five free credits after email verification, and each run uses one. Your edits are saved across the sign-in, so you do not lose them.
• Two models are on offer. jev-1.13 is a pinned build that will not change under you. jev-latest follows TypeSafe's newest release. Start with the pinned one while you tune thresholds.
• Failed runs cost nothing. If a run errors, nothing is charged.

Pasting a real ticket, answer or document matters more than any wording trick. Jev judges the state you give it; a toy example tells you very little about your production data.

Step 3: Write Questions Jev Can Answer Well

This is the step that decides whether your first result looks brilliant or random. TypeSafe's docs describe each question as a gut-check: the kind of judgement a knowledgeable person could make in a few seconds, given the right context.

Two details trip up almost everyone the first time:

• choice criteria are an object and score criteria are an array. Labels map to descriptions for choice; tiers are listed lowest first for score. Swap them and the request is rejected with a 400 before the model runs.
• Keep instructions short and specific. On jev-ai.org each question's instructions are capped at 1,000 characters, and one call can carry up to 20 questions.

Once a question works, resist the urge to make it clever. Adding a fourth question is nearly free; making one question do two jobs is where accuracy drops.

Step 4: Read the Probabilities, Not Just the Label

The winning label is the least interesting part of a Jev answer. The distribution is what lets your code decide how much to trust it.

• A choice split 0.97 / 0.02 / 0.01 is safe to automate.
• A split of 0.52 / 0.48 is not a decision — it is a case for a person or a slower model.
• A score of 1.89 on a four-tier scale is the expected value of the distribution, not a rounding error. Most of the weight sits on tier 2, some on tier 1. Store it as a decimal and sort on it.

Pick a threshold per question: act automatically above it, escalate below it. If a case a person finds obvious keeps coming back close to 50/50, the question is usually the problem. Narrow it, or add a line to the label descriptions, and run it again.

On speed, expect results in well under a second for ordinary payloads. In the six production-shaped requests we timed on September 25, 2026 against jev-1.13-20260917, the median round trip was about 310 ms from our machine, with one outlier near a second.

Step 5: Move the Same Request Into Code

When the answers look right, you do not rewrite anything. The body you built in the playground is the body your code sends; the playground's code panel shows it as cURL and TypeScript. Here is how that looks on each route.

On jev-ai.org

Create a key in Settings → API keys (it is shown once and starts with sk-glm5-), then send the request. This is the three-question example from our API docs:

curl https://jev-ai.org/api/v1/systemone/ \
  -H "Authorization: Bearer $JEV_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jev-1.13",
    "state": "We were billed $49 twice on September 3. I opened ticket #4417 four days ago and have had no reply, and our books close on Friday.",
    "questions": {
      "needs_human": {
        "type": "noul",
        "instructions": "Does this ticket need a human reply rather than an automated one?"
      },
      "queue": {
        "type": "choice",
        "instructions": "Route this ticket to the team that should own the first reply.",
        "criteria": {
          "billing": "Invoices, refunds, duplicate charges, plan changes, tax.",
          "technical": "API errors, integrations, outages, SDKs.",
          "sales": "Pre-purchase questions about plans, limits or trials."
        }
      },
      "anger": {
        "type": "score",
        "instructions": "Rate the customer'"'"'s frustration.",
        "criteria": ["Calm", "Mildly annoyed", "Frustrated", "Angry"]
      }
    }
  }'

Keep the trailing slash on /api/v1/systemone/; without it the site answers with a 308 redirect, which curl only follows with -L. The full request and response contract is in the Jev AI API reference. To score thousands of rows instead of one, upload a CSV to batch processing; each row uses one credit, or tokens once credits run out.

On TypeSafe directly

If you would rather call TypeSafe's own API — the most common reading of "how to use the typesafe.ai Jev model" — the shape is the same with a different host and key. From TypeSafe's quick start, get a key from the console dashboard, then:

curl -X POST https://api.typesafe.ai/v1/systemone \
  -H "Authorization: Bearer $TYPESAFE_API_KEY" \
  -H "Content-Type: application/json" \
  -d @- <<'EOF'
  {
    "state": "Hi, I've been trying to connect my Stripe account for 3 days and the integration keeps failing. I'm losing sales. Please help ASAP.",
    "model": "jev-latest",
    "questions": {
      "urgency": {
        "type": "noul",
        "instructions": "Does this message express urgency?"
      }
    }
  }
EOF

TypeSafe also publishes a Python SDK (pip install typesafe-sdk, Python 3.10 or later) that reads TYPESAFE_API_KEY from the environment and defaults to jev-latest, and an agent skill for coding assistants: in Claude Code, claude plugin marketplace add typesafe-ai/skills followed by claude plugin install typesafe@typesafe-ai.

On OpenRouter

OpenRouter lists the model as typesafe/jev-1.13, plus a rolling ~typesafe/jev-latest, at the same $0.042 per million input tokens with free output. Its Jev pages show a decisions call through the OpenRouter SDK rather than a chat completion, because the model does not generate text. If you already route every model through OpenRouter, this keeps Jev on the same bill.

Troubleshooting: Why Your First Jev Call Fails

Nearly every failed first call we see falls into one of these. Each has a fast check.

Rule of thumb: a failed call on jev-ai.org never costs anything, so when in doubt, fix one field and run it again rather than guessing at three.

If the problem is further upstream — you cannot sign up, you were told the service is full, or a waitlist email never arrived — that is an access issue rather than a request issue. Our checklist for "why am I not able to use Jev AI" is in the Jev waitlist guide.

FAQ

How do I try Jev AI without installing anything?

Open the playground in your browser, pick a scenario, paste your own text and run it. Editing is free without an account; running uses one of the five free credits new accounts receive.

Where can I try Jev AI for free?

On jev-ai.org, with the welcome credits and daily check-ins. TypeSafe's own console stopped giving free credits to new signups on September 27, 2026, and OpenRouter bills from the first call. We compare every free and paid option in is Jev AI free?

Can I use Jev like ChatGPT?

No. Jev never writes text, so there is no chat mode. If you need a summary, a draft or an explanation, use a generative model; use Jev for the decision that comes before or after it.

Does Jev work in Japanese or other languages?

It accepts any natural-language text, but TypeSafe states English is where accuracy is currently best. If you found this page searching for jev 使い方, the steps above are the same; just test on real examples in your language and watch the confidence values closely.

Should I use jev-1.13 or jev-latest?

Use the pinned jev-1.13 while you are tuning thresholds, so the answer distribution does not shift under you. Move to a new build on your own schedule. Every response includes model_version, so log it.

Can a coding agent use Jev for me?

Yes. TypeSafe publishes an agent skill for Claude Code and other coding agents, which teaches the agent the request format. The agent still calls the same API, with your key.

The Bottom Line

Using Jev comes down to one habit: stop asking it to write, and start asking it narrow questions with the answer shape declared.

• Four routes run the same model; the browser playground is the only one that is free to start today.
• Put the real text in the state, ask one idea per question, and read the distribution before trusting the label.
• The request you tune in the playground is the one your code sends — no rewrite.
• When a first call fails, the error code tells you which of eight common fixes applies.

Pick the one decision you would most like to stop paying an LLM to make. Open the playground, paste a real example, and see whether the probabilities agree with you.

Sources

• Quick start — TypeSafe AI docs — The console playground, API endpoint, cURL example, Python SDK and agent-skill install commands.
• Models — TypeSafe AI docs — Current rate limits, context budget, aliases, list price and language support.
• Introduction — TypeSafe AI docs — The three question types and the advice to keep each question atomic.
• TypeSafe: Jev 1.13 — OpenRouter — The OpenRouter model id, listed price and release date.

TypeSafe's access rules, rate limits and credits changed several times in September 2026; everything above was checked on September 30, 2026. jev-ai.org details come from our own documentation, and the latency figure is from our six measured calls.