Jev AI — Typed Decision API for Classification, Routing, and Scoring

What is Jev AI?

Jev AI is a decision API rather than a chat model. Instead of generating prose, it performs classification, routing, scoring, and safety checks: the caller supplies shared state — a ticket, a message, form fields, or agent context — plus typed questions, and receives typed answers with probabilities that application code can branch on, such as escalating when a result is marked urgent.

The service defines three question types. Choice handles classification into discrete options, Score returns a continuous value, and Noul answers yes/no questions. Up to eight questions can be asked per request in the playground, all sharing one state. Alongside the flagship Jev model (listed as typesafe/jev-1.13), the site catalogs 14 decision APIs, including Laya, Cloudflare Clef and Clef-Flash, Perplexity PPLX Decider V1 27B, Liquid D1, Together Tev1 4B, Inception Mercury Decide, Upstage Solar Decide, several Respan Span-01 variants, Kev 4B, and a multilingual Jev-Omni classifier covering text, images, audio, and video with open weights.

The brand is new: the site carries a 2026 copyright, blog posts dated September and October 2026, and an unusual disclaimer that it is not affiliated with, operated by, or endorsed by TypeSafe. No company name, country, or founding information is published.

How Jev AI Works

The workflow is give state, ask the question, let code act. A request goes to POST /v1/systemone with the shared state and typed question definitions; the response includes the answer plus probabilities and confidence so the system knows when it is unsure. Integration is available through a REST API and a Python SDK (typesafe_sdk), with API key management in the dashboard. A free online playground supports Jev runs without cost (Laya runs consume account credits), and an agent skill installs into Codex, Claude Code, or Cursor via npx skills add jev-ai/jev-agent-skill.

Documented integration patterns include routing requests to smaller or larger models, gating risky tool calls behind a judgment, triaging support or lead queues, and preserving key facts during context compaction. Interactive demos — flight search, a wiki link race, support triage, a grocery cart, invoice review, and a security alert — illustrate the pattern without writing code.

Main Features

• Typed question types — Choice for classification, Score for continuous values, and Noul for yes/no, all returning machine-usable output schemas.
• Probabilities and confidence — every answer carries uncertainty signals so callers can set thresholds and fallbacks.
• Batched questions per state — up to eight questions evaluated against the same state in one playground run.
• 14-model decision catalog — small and large deciders, including a multilingual Jev-Omni classifier over text, images, audio, and video (open weights).
• REST API and Python SDK — typesafe_sdk plus API key management for production integration.
• Agent-skill install — one command adds Jev to Codex, Claude Code, or Cursor workflows.

Pricing

The playground is free for Jev runs. Paid access uses one-time credit packs that do not auto-renew: Starter costs $10 for 100,000 credits with one workspace and 3 concurrent requests; Pro costs $100 for 1,000,000 credits with unlimited workspaces, 10 concurrent requests, fast lane, and parallel questions (marked recommended); Enterprise costs $1,000 for 11,000,000 credits (a 10 percent bonus) with unlimited concurrency, team workspaces, and priority support. Credits persist after purchase with no expiry stated. Laya playground runs consume account credits.

Common Use Cases

• Support and email triage — classifying incoming tickets by urgency before a human or a larger model spends time on them.
• Lead scoring — continuous scores over inbound leads with explicit confidence.
• Content moderation and guardrails — Noul checks gating risky tool calls or LLM outputs.
• Model routing — choosing which model handles a request before invoking it.
• Agent context management — deciding which facts survive compaction in long agent sessions.

Limitations

Jev currently accepts text, JSON objects, and arrays of text only — images, audio, and video are not supported in the main model, despite the separate Jev-Omni classifier. The site states accuracy on non-English input should be validated separately and that confidence is not the same as accuracy. The service judges but does not execute: the calling agent or application retains control of final actions. The free tier covers only the playground, so API usage requires purchased credits. As a new product with no published accuracy benchmarks, SLA, or company identity — and a model name referencing TypeSafe while disclaiming any affiliation — teams should pilot carefully before committing production traffic.

Summary

Jev AI is a focused piece of developer infrastructure: typed decisions with probabilities, delivered through an API, SDK, and agent integrations, priced in one-time credit packs that are cheap to trial. Its honest self-documentation and free playground are strengths; its newness, undisclosed operator, unbenchmarked accuracy, and text-only input are the reasons to treat it as an experiment first rather than default infrastructure.