Jev access: how to get a key and make your first call

Or skip both: the widget below calls the live API with no signup at all.

1. Get Jev access from TypeSafe: from the waitlist to the key dashboard

TypeSafe's current quick start directs you to its console playground and API-key dashboard. It does not describe a mandatory waitlist. Check your account for current access and billing; a Jagent key is a separate credential for this site.

Open the TypeSafe key dashboard, or try our browser playground before choosing a provider. Access instructions checked on October 4, 2026.

2. Get Jev access through a gateway, with no queue

Jev is already resold through gateways you may use. Their billing, rate limits and privacy terms apply, not TypeSafe's:

Gateway or direct: we measured both

Same state, same questions, 12 calls to each route on 2026-09-20, alternating so both saw the same network at the same minute:

Questions scale the same way on both routes (measured). The differences worth knowing, none of them documented anywhere we could find:

• Model id differs: the first-party API answers as jev-1.13.0, OpenRouter as typesafe/jev-1.13-20260917. Pin the one your transport actually returns.
• The first-party API reports input_tokens and output_tokens but NOT cost. The cost field in our measured data comes from OpenRouter, which adds it. Direct, you compute spend from tokens and the rate card yourself.
• GET /v1/models lists only jev-latest and jev-preview, but jev-1.13.0 is accepted as a model id anyway. jev-1.13 without the patch number is rejected — "Unknown model". Both aliases currently resolve to jev-1.13.0.

Response shapes match exactly. Score returns the same 0..n-1 weighted mean with a legend and index-keyed probabilities, and Noul returns { type, noul } with no confidence field on either path.

3. With Jev access in hand, make your first request

The endpoint is POST https://api.typesafe.ai/v1/systemone and the early-access model route is jev-latest. It is not the OpenAI chat-completions shape: you send a state plus questions, so you need TypeSafe's client rather than an OpenAI SDK pointed at a new base URL.

Install

pip install "typesafe-sdk>=0.5.7"

Python

import os
from typesafe_sdk import Choice, Noul, NoulCriteria, Score, TypeSafeClient

client = TypeSafeClient(api_key=os.environ["TYPESAFE_API_KEY"], timeout=120.0)

response = client.system_one(
    model="jev-latest",
    state="The payment failed twice and the customer is asking for a refund.",
    questions={
        "department": Choice(
            instructions="Which team should handle this?",
            criteria={
                "billing":   "Payments, invoicing, refunds",
                "technical": "Bugs, outages, integrations",
                "sales":     "Pricing, upgrades, new accounts",
            },
        ),
        "urgency": Noul(
            instructions="Does this convey urgency?",
            criteria=NoulCriteria(
                true="Explicitly time-sensitive",
                false="No urgency expressed",
            ),
        ),
        "frustration": Score(
            instructions="How frustrated is the customer?",
            criteria=["Calm", "Frustrated", "Very angry"],
        ),
    },
)

answers = response.answers
print(answers["department"].choice)      # -> "billing"
print(answers["department"].confidence)  # -> 0.0-1.0
print(answers["urgency"].noul)           # -> 0.0-1.0
print(answers["frustration"].score)      # -> 0.0-2.0, not a label (see .legend)
print(response.usage.input_tokens)

curl -X POST https://jev-agent.com/api/v1/systemone \
  -H "Authorization: Bearer $JEV_AGENT_KEY" \
  -H "content-type: application/json" \
  -d '{"state": "…", "questions": {"urgency": {"type":"noul","instructions":"Is this urgent?","criteria":{"true":"Time-sensitive","false":"Not urgent"}}}}'

JavaScript / TypeScript

A separate package, @typesafe-ai/sdk on npm. Take exact signatures from the JavaScript SDK reference rather than transliterating the Python above.

4. Things that will bite you

• It will not write anything. If your use case ends in a sentence, Jev is the wrong half of the pipeline. Pair it with an LLM.
• Options must be known at request time. Choice picks from the set you send, so a dynamic action space means building that list per call.
• Use the confidence value. A winning option at low confidence is a coin flip with extra steps — gate on a threshold and fall back to an LLM or a human.
• 64k tokens per request. The 64k budget covers the state plus every question in the request. A second, tighter 32k limit applies to the state plus the single longest question. Long documents need chunking or a retrieval step in front.
• Text only at launch. Text only — no image or audio input at launch

Get Jev access: four common questions