LiveKit Agents

Voice agents built with LiveKit Agents run as programmable participants on your servers, joining WebRTC sessions or phone calls and pairing perception with speech in real time. The framework's plugin ecosystem lets you swap STT, LLM, and TTS providers behind a stable API, installed as pip install livekit-agents with provider extras such as openai, deepgram, and cartesia. Semantic turn detection uses a transformer model to judge when a caller has finished speaking, cutting false interruptions, and the stack includes telephony through LiveKit's SIP integration, native MCP support, job scheduling and dispatch APIs, and a testing framework with LLM judges. The core framework is Python with a separate agents-js project for Node, and client SDKs cover all major platforms. Media transport runs on LiveKit's open source WebRTC server, self-hosted or via the paid LiveKit Cloud, so the whole media pipeline can stay on your infrastructure even though the speech and language providers are typically hosted APIs. Framework code is Apache-2.0, the turn detection models carry a LiveKit Model License, and the repo holds about 11.6k stars.