The Prompt API is a Chrome built-in AI surface that runs Gemini Nano on-device, exposed via the global window.LanguageModel. No network call, no API key, no per-token cost.

The flow is always: feature-detect → check availability → create → prompt. Pass the same options object to availability() and create() so the readiness check matches the session you build.

if (!("LanguageModel" in self)) { // fall back to a cloud model } else { const opts = { /* same for both calls */ }; const status = await LanguageModel.availability(opts); // "unavailable" | "downloadable" | "downloading" | "available" if (status !== "unavailable") { const session = await LanguageModel.create({ ...opts, monitor(m) { m.addEventListener("downloadprogress", e => console.log(`model ${Math.round(e.loaded * 100)}%`)); }, }); const answer = await session.prompt("Summarize: ..."); } }

The monitor hook matters because the first create() may trigger a multi-hundred-MB model download; surface progress to the user instead of appearing frozen. For token-by-token output use session.promptStreaming(text), which returns a ReadableStream.

Red flag: checking availability with one options object and creating the session with a different one — capability is keyed to the options you pass, so a mismatched pair means the check doesn't reflect the session you actually get.

I feature-detect, check availability, and create the session with one shared options object every time, with a download monitor wired in — so the UI never looks frozen during the one-time model download.