What it does

FastRTC wraps your Python handler in a Stream object and exposes it over WebRTC or WebSockets. You write echo(audio) or detection(image); the library handles the browser negotiation, media tracks, and turn-taking. It also bundles a Gradio UI, FastAPI mounting, and even a temporary phone number via fastphone().

The interesting bit

The ReplyOnPause handler is the quiet convenience: it detects when the user stops speaking, buffers the audio, and triggers your response logic automatically. For voice AI demos, that removes the usual “hold-to-talk” jank without you writing VAD code.

Key highlights

• One-liner Gradio UI: stream.ui.launch() for instant browser testing
• FastAPI integration: stream.mount(app) gives you WebRTC and WebSocket endpoints
• Built-in telephone bridge: stream.fastphone() spins up a callable number (Hugging Face token required)
• Optional VAD + TTS extras via pip install "fastrtc[vad,tts]"
• Cookbook spans Gemini, OpenAI, Claude, Whisper, YOLOv10, and Kyutai Moshi

Caveats

• The repo language tag says JavaScript, but the library is Python; likely a GitHub metadata quirk from the Gradio frontend components
• fastphone() and some demos depend on Hugging Face infrastructure; portability beyond that stack is unclear

Verdict

Worth a look if you’re prototyping voice or video AI and want to skip the WebRTC rabbit hole. If you already run custom SFUs or need fine-grained codec control, this abstraction will feel like a straitjacket.