Make your first API call
Call an available model without downloading weights or running a GPU. You need a Zils account and a terminal with Bash or zsh, and curl. To create an account, choose Get started on zils.ai and sign in with Google or an email link.
This developer quickstart routes a support request to one of three teams. You will get a structured answer your application can inspect; the API does not move the ticket. For the broader workflow, read how Zils fits into your application.
1. Create an API key
Sign in at zils.ai/train, open API keys, and create a named key. Save the secret when it is shown; it is displayed only once. Keep it in your server's secret storage, away from browser code and source control.
Key management uses your signed-in session. Prediction calls use the resulting Zils API key. These credentials are not interchangeable.
Requests are prepaid at $0.06 per million input tokens, so add credit to your account first; a request without credit returns HTTP 402. See pricing and limits.
Replace YOUR_ZILS_API_KEY in each curl command below with the API key you just created.
2. List available models
curl --fail-with-body 'https://training.zils.ai/decision/v1/models' \
-H 'Authorization: Bearer YOUR_ZILS_API_KEY'The following text example uses the shared alias zils-shared. Confirm it is available in your model list before continuing. The response identifies the release that actually answered; use an immutable ID from the catalog when you need to pin that release. For an image task, use the separate image request guide.
z3-mini is the base for new training jobs. Training profiles are different from callable model IDs; the model guide explains how to select an accepted adapter later.
3. Ask a typed question
curl --fail-with-body 'https://training.zils.ai/decision/v1/systemone' \
-H 'Authorization: Bearer YOUR_ZILS_API_KEY' \
-H 'Content-Type: application/json' \
-d '{
"model": "zils-shared",
"state": "Please send the invoice for my last order.",
"questions": {
"route": {
"type": "choice",
"instructions": "Choose the responsible team.",
"criteria": {
"billing": "Invoices and payments",
"technical": "Software troubleshooting",
"account": "Account access"
}
}
}
}'The response's answers.route contains the selected choice, per-option probabilities, and confidence. The top-level model identifies the resolved release; usage records the billed input tokens; output is not charged. Read the actual returned values rather than assuming this example establishes accuracy on your traffic.
The API also supports yes/no (noul) and score questions. How decisions work explains the fields and their meaning.
4. Connect the answer to your workflow
Inspect the selected team and its probability before taking an action. For example, billing 0.90, technical 0.06 and account 0.04 would favor billing. These are illustrative values, not an expected response from the request above.
Your application needs its own policy for routing, review and API failures. Choose thresholds using labeled examples from your task; a 0.90 probability does not guarantee a correct answer. Start by recording predictions alongside your current process before enabling automatic actions.
You have completed the API quickstart when the response contains a resolved model and a valid answers.route distribution. You have not yet established accuracy for your application.
Continue with your application
Install the Zils SDK or CLI for application integration. The API reference covers keys, bulk jobs, errors, capacity, and retention. Prepare a dataset covers task evaluation and optional training.
To order several options instead of selecting one answer, try experimental SDK ranking. The SDK uses pairwise ranking by default and offers optional grouped comparisons. The guide includes Python and TypeScript examples and explains relative scores, usage limits and the quality tradeoff.