1
Browse
Apps and showcases are narrow decisions: route a ticket, guard a tool call, check a citation.
Community site · not the official product
Play with Jev,
online and in your code.
Jev is TypeSafe AI’s System One model: you send program state and typed questions, it sends back a choice, a score, or a yes/no probability. This site is the sandbox around that — playbooks, a playground, and small API presets.
Browser playground for choice, score, and noul
Playbooks you can fork in one click
MCP tools that compile into official questions
Your key is forwarded, not stored
Get started
Three steps
Pick a playbook, run it against the real model, then keep the JSON.
2
Run
The playground builds a real systemone request. Paste a TypeSafe key and the answer comes back with probabilities.
3
Keep it
Copy the JSON into your app, or point an agent at the MCP tools and skills on this site.
from typesafe_sdk import Choice, Noul, Score, TypeSafeClient response = TypeSafeClient().system_one( state=ticket, questions={ "team": Choice(instructions="Which team should handle this?", criteria={...}), "urgent": Noul(instructions="Does this convey urgency?"), "heat": Score(instructions="How frustrated is the customer?", criteria=[...]), }, ) # response.answers["team"].choice, .probabilities, .confidence
What this site is
Community desk, official model
TypeSafe ships Jev. jevai.ai hosts the extras: a runner, preset workflows, and notes on how to wire the answers into ordinary code.
| TypeSafe / Jev | jevai.ai | |
|---|---|---|
| Model | Jev 1.13, trained for calibrated decisions | We call api.typesafe.ai with your key |
| Interface | SDK and POST /v1/systemone | Playground, six preset routes, and an MCP server |
| Output | choice, score, or noul — never free text | Same answers, drawn as bars you can read |
| Account | Early access and API keys | No account here. Bring the key you already have |
On the bench
What you can run today
Support triage
Team, urgency, and frustration in one parallel call.
Tool-call gate
Allow, confirm, review, or deny before a side effect.
Task route
Fast pass, deep review, split, or block.
Citation check
Does this evidence actually support the claim?
Invoice coding
One account, a split, or send it to review.
Context keep / drop
Decide which snippets stay verbatim.
FAQ
Straight answers
Is this the official Jev site?
No. Jev is made by TypeSafe AI. Official docs live at docs.typesafe.ai. jevai.ai is an independent playground and recipe site.
Does the playground spend money?
Yes, against your TypeSafe key. Jev’s published price is $0.042 per million input tokens, and output tokens are free. A short ticket is a fraction of a cent. We do not keep the key.
Can Jev write the reply to the customer?
No. It returns a typed decision. The email, the code, and the tool call stay with a generative model or with ordinary code.
How do I add a playbook?
Use the share form. Useful ones get a preset in the playground.
What is a noul?
TypeSafe’s word for a yes/no. The answer is a probability from 0 to 1, not a boolean. Near 1 is a strong yes, near 0 a strong no, near 0.5 means the model is split. It is not a “medium amount” of anything.
Why are the questions independent?
Every question in one request sees the same state and does not see the other answers. That is what makes a batch cheap and stable: adding “is this urgent?” does not change the team classification. If a later question truly needs an earlier answer, make a second request.
One ticket, three numbers
What you actually get back
Take a note that says payouts have failed for three days and sales are dropping. One call can ask three different shapes of question. None of them is a paragraph.
team = billing
Probabilities over the options you listed, summing to 1, plus a confidence that says how peaked that distribution is. A 0.84 / 0.15 / 0.01 split is a billing ticket that still mentions an integration.
urgent = 0.99
“Three days” and “losing sales” are time pressure. Your code decides the cutoff. 0.7 might page someone. 0.99 can skip the queue. The model does not pick the cutoff.
heat = 1.04
On a scale you wrote — calm, civil, angry — the score is the expected level. 1.04 on a 0–2 scale is “frustrated but civil,” with a little weight on the next rung. Confidence tells you whether that split is sharp.
Put the weights in a constant. When the policy changes, you edit the constant. You do not rewrite a prompt and hope the JSON shape survives.
Where it fits
Code, Jev, or a writer
| The job | Who should do it |
|---|---|
| Days overdue, a regex, a date comparison, a refund total | Ordinary code. Jev is a poor calculator and reads dates as text. |
| Which queue, how angry, does this quote support the claim | Jev. The answer set is closed and the input is messy language. |
| The email back to the customer, the patch, the incident write-up | A generative model, or a person. Feed it the decision and the evidence. Do not ask it to also invent the decision. |
| Send the refund, delete the row, call the tool | Your code, after the gate. Jev never runs the tool. |
What a call costs
The published card for Jev 1.13 is $0.042 per million input tokens. Output tokens are free. A few hundred tokens for a ticket is a fraction of a cent. The interesting number is how many of those calls you can put on a hot path: TypeSafe quotes roughly 70–500ms end to end, often near 100ms from the US West Coast, where the service runs today.
Questions in the same request run in parallel, so a triage with six questions is one round trip, not six. The state is ingested once. Extra questions cost the tokens in the question text, which is why the playground lets you ask the speculative ones up front and ignore the answers you do not need.
Early access is still a TypeSafe waitlist. This site does not mint keys. Paste the one from console.typesafe.ai and the playground forwards it.