Android

A Kotlin app with Jetpack Compose and Material 3 Expressive, on webrtc-sdk 150.7871.01. Android 7.0 (API 24) or newer.

Screenshot to add

Android in a 1:1 call, portrait: the other person's video fills the screen, your tile in a corner, and the floating toolbar visible at the bottom.

File: docs/public/media/android-call.png

Run the app configuration from Android Studio, or ./gradlew :app:installDebug in android/.

Where things are

All paths are under android/app/src/main/java/com/example/myapplication.

File	What it does
MainActivity.kt	Hosts the Compose lobby, starts CallActivity
CallActivity.kt	Hosts the call screen: permissions, pickers, MediaProjection, PiP
call/	BaseCallViewModel (shared state and media controls), CallViewModel (1:1), GroupCallViewModel (group)
webrtc/LocalMedia.kt	Factory, capturers, sources, tracks, effects, the 300 ms rule. Shared by both engines.
webrtc/PeerConnectionClient.kt	The 1:1 engine
webrtc/WebRtcPeer.kt	One PeerConnection, its data channel and cryptors
webrtc/SignalingSocket.kt	JSON over an OkHttp WebSocket, callbacks on the main thread. Used by both engines.
webrtc/sfu/GroupCallClient.kt	The group engine: publish and subscribe connections
webrtc/E2eeManager.kt	The FrameCryptor key provider
webrtc/effects/	GLES effects processor, MediaPipe segmenter and face tracker
ui/video/TextureViewRenderer.kt	The video view

How it does each part

• Switching sources. The camera and the screen each get a new track, put on the same sender with RtpSender.setTrack(). A file reuses the camera's VideoSource with another capturer. Switching video sources
• Screen sharing. MediaProjection with a foreground service, ScreenCaptureService.kt. It stays alive while you share another app, so it hears about rotation and resizes the virtual display.
• E2EE. FrameCryptor from webrtc-sdk with the shared options. End-to-end encryption
• Effects. A VideoProcessor on the camera's VideoSource. Everything runs in GLES shaders on the camera texture, and only a small copy is read back for the models. Backgrounds and effects
• Picture-in-picture. Enters by itself on the home gesture on Android 12+. Picture-in-picture

Video rendering

Video is drawn by TextureViewRenderer, a TextureView fed by an EglRenderer. Unlike SurfaceViewRenderer, a TextureView can be clipped, rounded and animated with the rest of the Compose tree, which the floating self view needs.

The renderer always fills its view. So the view is laid out at the frame's aspect ratio, just large enough to cover the screen, and "fit" scales it down with a graphicsLayer. Switching between fit and fill is then a GPU-only animation that never resizes the TextureView. iOS, macOS and Windows use the same idea.

Threads

Thread	Work
Main	UI, every PeerConnection call, socket callbacks
WebRTC signaling thread	PeerConnection.Observer callbacks, which hop to main
CaptureThread	Camera frames and the effects' GL work
SegmenterInference, FaceTracker	The MediaPipe models

Callbacks are dropped once a call has ended, because calling a disposed native PeerConnection crashes the process.

Notes

• CallViewModel survives rotation and owns the WebRTC client and the shared EglBase, so the call keeps running while the activity is recreated. CallActivity handles size changes itself, so in practice it isn't recreated at all.
• Camera2Session tags every frame with the device rotation, so the other side sees an upright picture whichever way the phone is held.
• Material 3 Expressive is only in the material3 1.5 alphas, pinned to 1.5.0-alpha18. See Troubleshooting before bumping Compose.
• The MediaPipe models (selfie_segmenter.tflite, face_landmarker.task) are stored uncompressed in the APK's assets.