Explainer · independent
What is Jev AI?
Jev AI is the community name for TypeSafe AI’s System One API. Instead of generating free text, it answers typed questions about a piece of state and returns constrained answers — Choice, Score, or Noul — that your code can branch on without a parsing step.
This page explains the model, the three question shapes, and what TypeSafe publishes about latency and pricing. Everything numeric below is quoted from first-party sources and attributed — this directory does not run its own benchmarks. You can also try Jev in the browser or read Jev vs Laya.
The core idea: a decision, not a document
A generative model answers with text, so the calling application inherits a parsing problem: the response must be validated, repaired, or retried when a bracket slips or a key changes. System One removes that step by constraining the answer space up front. You declare what answers are legal; the model returns one of them.
That single design choice cascades. Answers become loggable values. Thresholds become code, not prompt instructions. And because nothing is being generated token by token, the round trip lands in a much tighter window than a chat completion.
Jev vs LLM
The practical difference is not “smarter vs dumber” — it is output shape, whether your code must parse prose, latency magnitude, and where each tool belongs.
| Model | Output shape | Parsing required? | Latency magnitude | Best fit |
|---|---|---|---|---|
| Jev (System One) | Constrained value: Choice, Score, or Noul | No — branch on structured answers directly | Vendor-published ~70–500ms end to end (often ~100ms from US West Coast); not measured here | Routing, gates, triage, scoring — answers you already know the shape of |
| Generative LLM | Free-form text that continues until it stops | Yes — validate, repair, or retry when labels slip | Full chat-completion round trip; token cost stacks on hot paths | Prose, summarization, open-ended chat, non-enumerable answers |
For a same-primitive vendor comparison, see Jev vs Laya. For decision models against LLMs more broadly, see decision models vs LLMs.
The three question shapes
Patterns across this site are built from three primitives. Multiple questions can be asked against one piece of state in a single request.
Choice
Pick one option from a closed set you define. Useful for routing, classification, and component selection, where an invented label would break downstream code.
["refund", "bug", "billing", "other"]Score
Return a number inside a range you set. Useful for urgency, relevance, quality, or similarity ranking when you need to threshold in code.
urgency → 1…5Noul
A yes/no value, read here as a gate. Useful for safety checks, human-escalation flags, and circuit breakers that must be deterministic.
needs_human → true / falseWhen Jev fits — and when it does not
Good fit
- →You already know the shape of the answer — an enum, a number, or a boolean.
- →The decision sits in a hot path, a loop, or a per-item fan-out where generative latency multiplies.
- →Downstream code needs something it can branch on without parsing prose.
- →You want a decision you can log and audit as a value, not a paragraph.
Wrong tool
- ×You need prose, summarization, or a chat response — that is a generative model's job.
- ×The answer space is genuinely open-ended and you cannot enumerate candidates.
- ×You need the model to explain itself in sentences rather than return a value.
Latency and pricing, as published
Two numbers come up constantly when people evaluate Jev. Both are vendor-published figures, quoted here with attribution. For endpoint, SDKs, and a fuller pricing write-up, see the Jev API reference.
Verify before you rely on either number: typesafe.ai · docs.typesafe.ai
How to get a Jev API key
Keys come from TypeSafe directly — this directory does not issue or proxy them. Full endpoint, pricing, and SDK notes live on the Jev API page; the essentials are below.
https://api.typesafe.ai/v1/systemone@typesafe-ai/sdkPrefer not to write integration code first? try Jev in the browser.
Where to go next
Jev API reference
Endpoint, model naming, official list pricing, and JS/Python SDKs — for builders wiring System One.
try Jev in the browser
Paste text or upload a document, ask a typed question, and read the structured answer with measured latency.
Jev vs Laya
Hosted System One versus the open-weight Laya checkpoints — latency, license, and when to pick each, with every figure attributed.
Read the composition patterns
Routers, guards, triage, RAG verification, and CI triage — each written up as a concrete arrangement of the three primitives.
Browse tools by category
SDKs, integrations, agent tooling, browser agents, applications, and more from the community catalog.
Compare it with the alternatives
Decision models vs LLMs and open-weight replacements — every published figure attributed to its source.
Keep up with the ecosystem
New projects land in the community list first — that is the source this directory syncs from.
Frequently asked questions
What is Jev AI?
+
Jev AI is the community name for TypeSafe AI's System One API. Instead of generating free text, it answers typed questions about a piece of state and returns constrained answers — one of your choices, a score in a range you define, or a yes/no value. TypeSafe ships the model and infrastructure; this site is an independent directory of tools built around it.
Is Jev a large language model that writes prose?
+
No. A generative LLM continues text, so downstream code has to parse it. System One is asked a question with a constrained answer space, which means the caller receives structured values directly. That is the main practical difference, and it is why applications can branch on the answer without a JSON-repair step.
What are Choice, Score, and Noul?
+
They are the three question shapes referenced throughout this directory. Choice picks from a closed set of options you provide. Score returns a number within a range you set. Noul is a two-option decision — used here to describe yes/no style gates such as "needs human" or "block this request". Multiple questions can be asked against one state in a single call.
How much does Jev cost?
+
Official list pricing as published by TypeSafe in September 2026 is $0.042 per 1M input tokens with output tokens free. This directory quotes that figure as-is and does not run its own pricing tests; check https://typesafe.ai for current rates before budgeting.
How fast is Jev?
+
TypeSafe publishes an end-to-end latency range of roughly 70–500ms, with many calls landing near 100ms from the US West Coast. Treat that as a vendor-published range rather than a benchmark. The per-pattern millisecond figures shown on this site are illustrative labels for the patterns, not measurements taken here.
How do I get a Jev API key?
+
Get a key from TypeSafe directly. The official JavaScript SDK is published as @typesafe-ai/sdk and the HTTP endpoint is api.typesafe.ai/v1/systemone. The community Discord is the usual place to ask about access, limits, and previews.
Can I try Jev without writing integration code?
+
Yes. This site runs a browser playground where you paste text or upload a document, ask a typed Yes/No, Choice, or Score question, and see the structured answer. Several entries in the demos and playgrounds category are also hosted and runnable in the browser.
Is jevaitools.com affiliated with TypeSafe AI?
+
No. It is an independent third-party directory. The catalog is sourced from the community-maintained awesome-jev list, and official resources are linked so you can verify anything quoted here at the source.