expo-ai-kit weekly downloads statistics

expo-ai-kit

On-device AI for React Native. Generate text, transcribe speech, work with images, and search by meaning. Supports Expo apps and React Native apps with Expo modules installed. Inference runs on the device, with no API key.

Documentation · Examples · npm · Changelog

Install

npx expo install expo-ai-kit

All features are off by default. Nothing native is compiled into your app until you turn a feature on in your app config and make a native build. To use text generation:

{
  "expo": {
    "plugins": [["expo-ai-kit", { "llm": true }]]
  }
}

Requires Expo SDK 54+ (or compatible Expo modules in an existing React Native app) and a native development or production build. Expo Go is not supported. Follow Get Started for native setup and a working app example. Device and OS requirements vary by feature; see platform support.

First request

With llm enabled, on a device with a supported built-in text model:

import { isAvailable, prepareBuiltInModel, sendMessage } from 'expo-ai-kit';

async function ask(question: string) {
  if (!(await isAvailable())) {
    return 'The built-in text model is unavailable on this device.';
  }
  await prepareBuiltInModel();
  const { text } = await sendMessage([{ role: 'user', content: question }]);
  return text;
}

const answer = await ask('Explain gravity in one sentence.');

The built-in text model requires Apple Intelligence on iOS 26+ or a supported Android device. Android may download its model during preparation. You can also download a model such as Gemma, Qwen, or Phi.

LLM

Generate and stream text, return structured JSON, or let the model call your functions. Pass the full message history on each call. One generation runs at a time. Enable the llm option and rebuild.

Speech

Transcribe microphone audio or recorded files with streamTranscription and transcribe. Enable the speech plugin option and rebuild. iOS requires 26+; Android requires 12+ and processes file audio at real-time rate.

Vision

Remove backgrounds, label images, and read text with removeBackground, labelImage, and recognizeText. Find faces with detectFaces:

import { detectFaces } from 'expo-ai-kit';

const { width, height, faces } = await detectFaces({ uri: photo.uri });
// faces[0].bounds (normalized 0–1) and faces[0].pixelBounds, largest face first

Enable vision on Android and rebuild, or name only the features you use, for example { "vision": ["face-detection"] }, so the app ships only that detector. Face detection uses a bundled model and needs no download; rules such as "exactly one face" are a few lines in your app.

Embeddings

Semantic search and retrieval with embed, chunkText, and createVectorStore. iOS uses the system embedding model; Android requires the androidEmbeddings option and an explicit model download.

Build options

Every option is off by default. Each one compiles a single feature into the app; leave it off and that feature's native code, models, and permissions stay out of the build, and its functions throw a typed *_NOT_ENABLED error. Changing an option requires a new native build.

Option	Enables	Adds to the app
llm	Text generation, JSON output, tool calling, model downloads, the AI SDK provider	LiteRT-LM runtime (about 30 MB of iOS arm64 code, 21 MB on Android); Android minSdkVersion 26
speech	transcribe, streamTranscription	Android ML Kit speech and microphone permission; iOS microphone usage string
vision	Android removeBackground, labelImage, recognizeText, detectFaces	true adds every ML Kit vision client and bundled model; an array such as ["face-detection"] adds only those features. iOS needs no option
androidEmbeddings	Android embed	MediaPipe TextEmbedder (about 25 MB); the model downloads at runtime

The config reference covers bare React Native setup and every option in detail.

More

• Vercel AI SDK: use expo-ai-kit/ai with the AI SDK.
• API reference: functions, options, results, and typed errors.
• Troubleshooting and GitHub issues.
• llms.txt: documentation index and essential contracts for coding agents.

Zero runtime JavaScript dependencies. MIT © Said Kaban.