What is Jev, and when should you use its API?
Jev is a decision model from TypeSafe. Give it the context for a bounded question and inspect a typed answer. In this workspace, a practical first task is assigning a support ticket to billing, technical support or account access. Your application decides what happens after the classification.
TypeSafe calls this a System One model. A general chat model can write an explanation or plan a sequence of tasks; this API is useful when your code needs an explicit category, score or yes/no-style judgment. It does not browse, execute a tool or replace your agent’s planner. Choose based on tests with your own examples, rather than assuming one model is always faster, cheaper or more accurate.
Try a decision before creating an API key
Open the playground, choose a support, lead or feedback example, and edit the message and labels. Viewing the labeled sample is illustrative and does not call a model. Running a live decision is a separate action, subject to the displayed trial limits and service availability.
Check whether the output answers the business question you actually have. Use anonymized examples, include enough context to distinguish the labels, and keep a separate set of examples for evaluation. A convincing single result is not an accuracy benchmark.
Get your Jev API key
Sign in to this workspace, choose a plan with active API access, and create a named key in your dashboard. Copy the secret when it is shown and store it on your server. A key from TypeSafe, OpenRouter or another community website is not a Jev API Pro account key.
Send your key as an Authorization: Bearer header. A website login cookie alone does not authorize the REST endpoint. Keep secrets out of browser JavaScript, public repositories, prompts and Skill files. Revoke an unused or exposed key in the dashboard.
Send the quickstart request and read the response
The curl request above calls POST https://jevapi.pro/api/v1/decisions. Replace YOUR_API_KEY with your server-side key. Supply text, instructions, a list of distinct labels and a confidence threshold. Change the Idempotency-Key for each new decision; reuse it only when retrieving or retrying the same unchanged request.
A new successful decision returns HTTP 201. Read choice as the selected label and needsReview as the review flag. Confidence can be null when the provider does not supply it; do not turn a missing value into certainty. Keep id for support and tracing. The full reference describes field limits, multiple questions and response shapes.
Connect Jev to an agent through REST
Keep your agent’s main model and tool permissions in place. At a routing boundary, have your server call the same REST endpoint with a concise task description, explicit instructions and permitted outcomes such as proceed or review. Validate the returned choice against those outcomes before choosing the next step.
If needsReview is true, confidence is absent, or the request fails, pause the automatic action and use your review path. A model prediction never grants permission to issue a refund, send a message or delete data. Enforce those rules in your application. This workspace documents REST access; an MCP URL or jev_* tool name from a different website is not an endpoint here.
{
"text": "Task: read a public API guide and summarize its supported request fields. No account changes or external messages are requested.",
"instructions": "Choose review if the task is unclear or could change external state. Otherwise choose proceed. This classification never grants permission to execute a tool.",
"labels": [
"proceed",
"review"
],
"threshold": 0.9
}Recover from errors without duplicate work
Do not retry every failure. Fix invalid inputs for a 400 response, check the key and its API access for 401/403, and inspect credits for 402. For 429, respect Retry-After when supplied and cap your retries. A provider outage should reach a visible fallback, not an endless loop.
Replaying a completed request with the same key and payload returns the stored result without a second decision charge. A different payload with that key is rejected with 409. A recorded failed request also returns 409 on replay; after resolving its cause, use a new key for a deliberate new attempt. A timeout with an unknown outcome needs a same-key lookup first.
Jev API pricing and the next step
Jev API Pro sells prepaid access periods with decision credits, rather than the provider’s direct token pricing. Plans include web tools and API access; the current prices, included credits, access duration and checkout terms are on the pricing page. Buying here does not add credit to an OpenRouter or TypeSafe account.
Before paying, confirm that the playground handles your task. Before production, test ambiguous inputs, missing confidence, an expired key and provider failures. Start with suggestions that a person can check, then automate only the decisions your evaluation supports.
Jev API pricing and included credits · API security and data handling