Call an AI model from a browser app without exposing an API key

import { createSeige } from "@seige_sdk/sdk";

const seige = createSeige({ appName: "Trip planner" });

const { data, error } = await seige.ai.generate({
  prompt: "Suggest a packing list for this trip.",
  input: { city: "Lisbon", nights: 3, month: "October" },
});
if (error) console.warn(error.code, error.hint);
else console.log(data.text, data.usage);

What happens on the first call

The first seige.ai.generate() call opens a Seige window that asks the user to allow the app to generate AI answers. After they approve, the call continues and returns the answer. Later calls do not ask again in that browser. To ask before the feature is used, call seige.request({ access: "ai" }).

Structured answers

Pass a schema with a parse method, or a compatible schema library object. The SDK asks the model for JSON, parses the answer and validates it. The validated value is returned as data.value. Invalid JSON or a rejected value is returned as an error. Describe the expected fields in the prompt.

const { data } = await seige.ai.generate({
  prompt: "Return {\"items\": string[]} with up to eight items to pack.",
  input: trip,
  schema: { parse: (value) => ({ items: value.items.map(String) }) },
});

Cancelling and limits

Pass an AbortSignal as signal to cancel a request.
The prompt is limited to 16,000 bytes and the input to 100,000 bytes. Output defaults to 2,048 tokens, with a maximum of 4,096.
An app can make 20 requests per hour, within 100 requests per hour for all apps on the account, and one request per account runs at a time. seige.ai.usage() reports the current window.
When a request's estimated cost exceeds available AI credit, calls return ALLOWANCE_EXCEEDED. Pending requests reserve credit until settlement. AI credit has no automatic hourly reset. Do not retry automatically.

What the model receives

seige.ai.generate() is a standard model endpoint, not an agent. It has no tools and no access to the user's account. It receives only the prompt and input that your app sends, and your app can only send data the user gave it. AI inputs are not end-to-end encrypted: Seige and the model host process them in readable form. The approval window tells the user this before they allow access.

When this fits

This approach suits apps where each signed-in user asks for AI help with their own data. It does not suit anonymous visitors, because every request needs a signed-in Seige user who approved AI access, or apps that need tools, long agent loops or a specific model provider.

This guide is also available as Markdown for coding agents.