LiveKit Vision Demo

March 10, 2026 · View on GitHub

Warning

This example is outdated. While it still works great, we've now added built-in vision support to a whole set of new frontend starter apps for every platform and live video is easy to add to the Python agent starter repository for the latest example.

LiveKit Vision Demo

This LiveKit sample app shows a voice AI assistant with realtime audio and video input.

It contains a native iOS frontend, built on LiveKit's Swift SDK, and a backend agent, built on LiveKit's Python Agents framework and the Gemini Live API.

Features

Real-time Video & Audio

  • 📱 Front and back camera support
  • 🎙️ Natural voice conversations
  • 🖥️ Live screen sharing

Background Support

  • 🔄 Continues running while using other apps
  • 💬 Voice conversations in background
  • 👀 Screen monitoring while multitasking

The assistant can observe and interact with you seamlessly, whether you're actively using the app or working on other tasks.

Agent Architecture

The backend agent is built on the MultimodalAgent class hooked up to the Gemini Live API.

Video frames are sampled at 1 frame per second while the user speaks, and 0.3 frames per second otherwise. Images are sent as JPEG at 1024x1024 max size. For more information on video input, see the LiveKit Agents vision docs.

Running Locally

This project is meant to be a starting point for your own project, and is easy to run locally.

Running the Agent

Prerequisites

Setup

Put your LIVEKIT_URL, LIVEKIT_API_KEY, LIVEKIT_API_SECRET, GOOGLE_API_KEY into a file called agent/.env.

Then install dependencies

cd agent
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Finally, run the agent with:

python main.py dev

Using the Agents Playground

This project is fully compatible with LiveKit's Agents Playground, so you can easily test the agent in your browser without having to build the iOS app. Just go to the playground, pick your cloud project, and connect! There is a checkbox to "Enable camera" if you wish to share your camera feed with the agent.

Running the iOS App

This project includes a sample iOS app that you can build yourself.

Prerequisites

Setup

  1. Open swift-frontend/VisionDemo/VisionDemo.xcodeproj in Xcode.
  2. Create a file swift-frontend/VisionDemo/Resources/Secrets.xcconfig with LK_SANDBOX_TOKEN_SERVER_ID= and your token server's unique ID.
  3. Edit the bundle identifier for the VisionDemo target to a suitable values for your own use.
  4. Edit the bundle identifier for the BroadcastExtension to <your-bundle-identifier>.broadcast.
  5. Create a new App Group called group.<your-bundle-identifier> and select it in the "Signing & Capabilities" section of the VisionDemo target.
  6. Build and run the app on your device.

Self-Hosted Options

This project is built with the LiveKit Cloud token server (enable from your project's Options on the Settings page) to make token generation easy. If you want to self-host or run a local LiveKit instance, you'll need to modify swift-frontend/VisionDemo/Services/TokenService.swift file to fetch your token from your own server and remove the noise-cancellation plugin from the agent (enhanced noise cancellation is a LiveKit Cloud feature).