The Prompt API, in every browser

Write your AI feature for the web standard, the Prompt API (LanguageModel), and it works for more of your visitors. Where the browser has a built-in model that can run, that model answers. Elsewhere, @leuria/prompt-api provides LanguageModel, and the visitor's own AI answers through Leuria: ChatGPT, Claude, Mistral or a model on their computer. This covers browsers without a built-in model, and computers that can't run one.

npm install @leuria/prompt-api

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

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

const availability = await LanguageModel.availability()
if (availability !== "unavailable") showSummarizeButton()

// In the button's click handler: the first time, the visitor approves your site in Leuria.
summarizeButton.onclick = async () => {
  const session = await LanguageModel.create({
    initialPrompts: [{ role: "system", content: "You summarize notes in three bullet points." }],
  })
  for await (const chunk of session.promptStreaming(note.text)) output.append(chunk)
}

The rest of your code is plain Prompt API: nothing else to learn, and nothing to change if you remove the package later.

Without a bundler:

<script type="module">
  import { installPromptAPI } from "https://cdn.jsdelivr.net/npm/@leuria/prompt-api/+esm"
  await installPromptAPI({ app: "My notes" })
</script>

Who answers

installPromptAPI() leaves the browser's own LanguageModel in place wherever it can run, even if its model must download first. It steps in only where there is none, or where availability() says "unavailable". The when option changes this:

when	Leuria provides LanguageModel
"unavailable" (default)	Where the browser has none, or has one that can't run on this device
"missing"	Only where the browser has none
"always"	Always, in place of the browser's

It resolves with "browser" or "leuria", if you want to know. To get the class without touching the global, use createLanguageModel(options).

How the API behaves with Leuria

Prompt API	With Leuria
availability()	"available" once your site is connected, "downloadable" before, "unavailable" when the visitor has no Leuria
create()	Connecting is the "download": call it from a click the first time. Leuria asks the visitor to approve your site, and monitor gets downloadprogress 0, then 1. Without a click, it rejects with NotAllowedError
prompt(), promptStreaming()	Each LanguageModel is one session of the visitor's AI, which keeps the context: a prompt sends only what's new. Prompts run one after the other
initialPrompts	The system message becomes the AI's instructions; the others are the conversation so far
tools	Your tools' execute functions run in your page, called by the visitor's AI with the arguments object
responseConstraint	A JSON Schema or a RegExp. The AI gets it as an instruction and the answer is checked; an answer that doesn't match goes back to the AI to fix, twice at most, and then the prompt rejects
append(), clone(), destroy()	Supported. destroy() ends the AI session, and so does leaving the page
signal	Cancels the prompt at once; the AI stops its turn
Images	Accepted when the visitor's AI takes them (expectedInputs: [{ type: "image" }])
Audio	Not supported: availability() says "unavailable"
contextUsage, contextWindow, measureContextUsage()	Estimates, about 4 characters per token. The visitor's AI manages its own context, so contextoverflow never fires
topK, temperature, samplingMode	Accepted and ignored: the visitor's AI uses its own settings

The first prompt of a session takes a few seconds while the visitor's AI starts; later ones answer at once. Create the session when the feature opens, not at the last moment.

Options

installPromptAPI(options) and createLanguageModel(options) take:

Option	Default	What
app	none	The name Leuria shows when the visitor connects your site
needs	none	What your features need (tools, images, effort, context), so Leuria recommends an AI and model that fit. See Providers
skills	none	Skills that guide the visitor's AI on your site. See Skills
system	built in	Instructions for the AI when your page gives no system message
maxSteps	20	Most tool calls per prompt
retries	2	How many times an answer that misses the responseConstraint goes back to the AI
provider	leuria()	Another provider from @leuria/client, e.g. server({ url })

The Prompt API or the SDK?

Both reach the same AI. Pick by how you want to write your feature:

• The Prompt API (@leuria/prompt-api): your code is the web standard, and the browser's own model answers where it can. Good for features already written for Chrome's built-in AI, and for code you want to keep free of any library.
• The SDK (@leuria/client, see the Quickstart): one client over several AIs in the order you choose, including your server as a fallback. It adds conversations that survive a change of AI, tool middleware, turn context, structured output with validation, embeddings, and the Connect UI.

A page that uses both doesn't count the visitor's AI twice: promptAPI() in the SDK ignores the LanguageModel that @leuria/prompt-api installs.