Jev: the AI that decides instead of talking

Where to get Jev

How to use it

Pick the route that fits you. All of them hit the same model.

A. No code: the Playground

1. Open the Playground and log in.
2. Paste any text as the state, for example a customer complaint.
3. Add a question, like "Does this message express urgency?" as a Noul.
4. Add more questions. Mix all three types and see every result at once.

B. Python SDK

1. Install it (needs Python 3.10 or newer): pip install typesafe-sdk
2. Set TYPESAFE_API_KEY in your environment.
3. Send a state plus questions. The SDK defaults to jev-latest.
4. JavaScript or TypeScript instead: npm install @typesafe-ai/sdk (Node 20+).

Copyfrom typesafe_sdk import Choice, Noul, Score, TypeSafeClient

client = TypeSafeClient()   # reads TYPESAFE_API_KEY

ticket = "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."

response = client.system_one(
    state=ticket,
    questions={
        "department": Choice(
            instructions="Which team should handle this",
            criteria={
                "billing": "Payment or subscription issues",
                "technical": "Bugs or integration problems",
                "sales": "Pricing or account questions",
            },
        ),
        "frustration": Score(
            instructions="How frustrated the customer appears",
            criteria=[
                "Calm, just stating facts",
                "Frustrated but civil",
                "Very angry, strong language",
            ],
        ),
        "is_urgent": Noul(
            instructions="The message conveys urgency or time-sensitivity",
        ),
    },
)

print(response.answers["department"].choice)  # "technical"
print(response.answers["frustration"].score)  # 1.0
print(response.answers["is_urgent"].noul)     # 1.0

Example from TypeSafe's official quickstart.

What comes back

# department   technical   (technical 0.85, billing 0.15, sales 0.00, confidence 0.78)
# frustration  1.0         ("Frustrated but civil", confidence 1.0)
# is_urgent    1.0

Copycurl -X POST https://api.typesafe.ai/v1/systemone \
  -H "Authorization: Bearer $TYPESAFE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "state": "I was charged twice. Please fix this ASAP.",
    "model": "jev-latest",
    "questions": {
      "urgency": {
        "type": "noul",
        "instructions": "Does this message express urgency?"
      }
    }
  }'

Copyclaude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai

# other agents (Codex etc.)
npx skills add typesafe-ai/skills --skill typesafe-ai

Best use cases

Jev fits when the same decision repeats a lot and the possible answers are known in advance. Think of it as a smart if statement for judgments a normal if can't make.

5 beginner projects to build with it

• Support-ticket routerChoice for the team, Score for frustration, Noul for urgency.
• Internship / job screenerScore each application on 3 to 4 skills, then combine with your own weights.
• Comment and DM sorterTag incoming messages by intent and flag spam or urgent ones.
• Research-paper triageNoul "is this an RCT?" over a batch of abstracts.
• LLM guardrail wrapperCheck every user message for jailbreaks before it hits Claude or GPT.

Jev vs Claude, GPT, Gemini

TypeSafe's own benchmark

Four workflows: security incidents, agent traces, invoices, customer service. "Agreement" means how closely a model matches the average answer of two top frontier models. It is not ground truth.

Best pattern, from TypeSafe's own docs: let Jev classify and route at high volume, let plain code handle what it can, and hand only the hard minority to a big LLM.

Where it breaks

TypeSafe publishes its own list of weak spots for jev-1.13. Read this before you ship anything.

Sources

• TypeSafe launch post (company, 15 Sep 2026)
• TypeSafe docs: Quick start, Models, Jev 1.13 jaggedness, Use-case map
• DataCamp: Jev explained (independent write-up, 16 Sep 2026)
• DEV Community: practical guide and launch-week builds
• LangChain: building a harness with Jev and Vercel: what is Jev

Compiled 21 Sep 2026 from public sources. Speed, price and benchmark figures are TypeSafe's own unless stated. Details may change quickly, so check the official docs before building.