Getting started

One state in, typed questions out, decisions in 70 to 500 milliseconds. This page gets you from zero to a first call, then to a system that will not embarrass you in production.

The 60-second example

A support ticket, three questions, one request. The threshold is ordinary application code.

curl -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 for my subscription. Please fix this ASAP.",
    "model": "jev-latest",
    "questions": {
      "team": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {
          "billing": "Payments and refunds",
          "technical": "Bugs and integrations",
          "other": "None of the above"
        }
      },
      "is_urgent": {
        "type": "noul",
        "instructions": "Does this message convey urgency?"
      }
    }
  }'

The response carries a probability for every option plus a confidence value. Your code compares the selected probability to a threshold you chose, and decides. That split is the whole architecture.

Every access route

Pick the one that matches your stack. Two of these skip the waitlist entirely.

Route	How in	Notes
TypeSafe direct	Official SDK with a TypeSafe API key	The reference contract. Cache this page's shape in your head: `state` plus typed `questions`. Signups reopened on 28 September after a week of suspension; the $5 starter credit is currently withdrawn over abuse, so budget for real usage.
Vercel AI Gateway no waitlist	`typesafe-ai/jev` through the AI SDK evaluate path, an HTTP API at /v1/evaluate, or an existing TypeSafe client	The fastest way in while the waitlist moves. Boolean maps to Noul; the interface differs from systemOne.
OpenRouter no waitlist	Decisions API with `typesafe/jev-1.13`	Use the decisions endpoint shape. Do not send these questions to a chat-completions API, it will not work.
Cloudflare Workers AI	`typesafe/jev` through a Workers AI binding	Check Cloudflare's request shape and third-party model labeling.
Netlify AI Gateway	Official JS SDK from a Netlify Function	Server-side path, not a browser key.

SDKs

Official clients for JavaScript and Python. The community covers almost everything else, including Rust, Go, Ruby, Java, Kotlin, Swift, .NET, Elixir, Scala, PHP, R, and C++. Several are now near-duplicates of each other, so check whether an official client or a well-maintained package already covers your language.

JavaScriptnpm install @typesafe-ai/sdk, typed answers inferred from your question set

Pythonpip install typesafe-sdk, sync and async clients

Adaptersystem-one-adapter-python runs OpenAI and Anthropic models over the same interface, for honest side-by-side comparisons

Rusttypesafe-rs, s1-rs, and several others in the directory

Six design rules

Distilled from the official docs, the cookbooks, and the independent studies. Each one has a scar behind it.

Atomic questionsOne gut-check judgment per question. Decompose compound asks, then combine the answers in your own formula.

Filter firstSend only the material the question needs. Accuracy falls as state fills with irrelevant detail, and cost follows state size.

Criteria are instructionsContradictory instructions and criteria confuse the model. Treat the criteria text as part of the question, not fine print.

Always an escape hatchAn unknown or none option in every Choice. Forced binaries pick the least-wrong answer.

Code owns everything elseThresholds, math, dates, side effects, retries, and every irreversible action. The model supplies judgments, not control flow.

Measure before automatingLabeled set first, threshold second, automation third. The rigor page has the checklist and the failure studies.

Getting started

The 60-second example

Every access route

SDKs

Six design rules

Where to go next

The Playground

The cookbooks

Official patterns

The usecases