**LiveKit Agents** is a Python framework for building realtime voice AI — programmable participants that run on servers and can hear, see, and speak in a live session. Rather than shipping its own speech models, it defines the **session loop** around them: audio arrives over WebRTC, a voice activity detector and turn detector decide when the user has finished speaking, STT transcribes, an LLM reasons and calls tools, and TTS speaks the answer back with barge-in handled. The integration surface is deliberately open. Any combination of **STT, LLM, TTS, or a realtime speech-to-speech API** can be composed inside an `AgentSession`, with plugins for Deepgram, OpenAI, Cartesia, and many other providers — or LiveKit Inference, a unified hosted API, when you would rather not manage provider keys. Swapping a TTS vendor is a one-line change, which is the practical argument for the abstraction. The piece hardest to build yourself is **semantic turn detection**. Fixed silence thresholds either cut users off mid-thought or leave dead air; LiveKit runs a transformer model over the conversation to judge whether a pause is a real end-of-turn, which is where most of the perceived latency and interruption quality in a voice agent lives. Alongside it sit **integrated job scheduling** with dispatch APIs to route users to agents, and **telephony** via LiveKit's SIP stack, so an agent can place and receive ordinary phone calls without a separate bridge. The framework also ships things voice projects usually improvise: a **built-in test framework** with LLM judges for asserting an agent behaves as intended across turns, native **MCP support** that registers tools from an MCP server in one line, and RPC and data APIs for exchanging structured state with clients. Client SDKs cover the major platforms, and a companion **AgentsJS** library serves JavaScript and TypeScript teams. The caveat is architectural rather than a defect: the stack is designed around LiveKit's own WebRTC media server. That server is open source and self-hostable, so nothing here is locked behind the hosted product, but adopting Agents means adopting LiveKit's media plane, and the smoothest path — LiveKit Inference and managed dispatch — runs through LiveKit Cloud. Both the framework and the server are **Apache-2.0**, and the project releases frequently, with **livekit-agents 1.6.9 on 2026-08-07**.