@leuria/prompt-api

The Prompt API (LanguageModel) where the browser can't run it, served by the visitor's own AI through Leuria. Write your feature for the web standard once: where the browser has a built-in model that runs, it answers; elsewhere, the visitor's AI does (ChatGPT, Claude, Mistral or a local model), in any browser.

import { installPromptAPI } from "@leuria/prompt-api"

await installPromptAPI({ app: "My notes" })

if ((await LanguageModel.availability()) !== "unavailable") {
  // From a click the first time: the visitor approves your site in Leuria.
  const session = await LanguageModel.create({ initialPrompts: [{ role: "system", content: "You are concise." }] })
  console.log(await session.prompt("Summarize this note: …"))
}

installPromptAPI keeps the browser's own LanguageModel wherever it can run (when: "unavailable", the default). Use when: "missing" to step in only where the browser has none, or createLanguageModel() to get the class without touching the global.

How the API maps onto Leuria

| Prompt API | With Leuria | | --- | --- | | availability() | "available" once the site is connected, "downloadable" before (connecting is the "download"), "unavailable" without Leuria | | create() | From a click the first time: Leuria asks the visitor to approve the site. downloadprogress reports 0, then 1 | | prompt(), promptStreaming() | One session of the visitor's AI per LanguageModel; each prompt sends only what's new | | tools | Run in your page, called by the visitor's AI | | responseConstraint | JSON Schema or RegExp, given to the AI and checked; a mismatch goes back to the AI to fix | | append(), clone(), destroy(), signal | Supported | | Images | Supported when the visitor's AI takes them. Audio is not supported | | contextUsage, measureContextUsage() | Estimates (about 4 characters per token) | | topK, temperature, samplingMode | Accepted and ignored: the visitor's AI sets its own |

Docs: leuria.dev. Apache-2.0.