Skip to main content
Back

LiveKit Agents Hits 13K Stars Building Voice AI Without the Lock-in

LiveKit's open-source Python framework for realtime voice agents crossed 13K stars, packing telephony, MCP, and semantic turn detection into one stack.

Repo, Full Access
  • LiveKit Agents crossed 13K stars, gaining ~1,000 in a week for its realtime voice AI framework.
  • Bundles STT, LLM, TTS pipeline with semantic turn detection using a transformer model.
  • Native SIP telephony lets agents place and receive real phone calls without extra services.
  • MCP support integrates external tool servers with a single line of code.
  • Powers ChatGPT Advanced Voice; supports 50+ models via Inference, 200+ via plugins.
  • Apache 2.0, self-hostable, with built-in pytest-style test framework using LLM judges.

Building a voice agent that actually feels natural to talk to is deceptively hard. You need speech-to-text, an LLM, text-to-speech, a way to detect when the user stops talking, interrupt handling, and some kind of transport layer that can move audio around without adding latency. LiveKit Agents bundles all of that into a single Python framework, and it just crossed 13,000 GitHub stars with a thousand added in the last week alone.

The pitch is straightforward. The framework is designed for building realtime, programmable participants that run on servers, used to create conversational, multi-modal voice agents that can see, hear, and understand. It runs on top of LiveKit's WebRTC media server, which is the same infrastructure OpenAI uses for ChatGPT's voice mode, along with Character.ai and thousands of smaller vendors.

LiveKit Agents framework banner with sample Python code

What's in the box

The framework handles the plumbing you would otherwise stitch together yourself. It includes components for the core challenges of realtime voice AI, such as streaming audio through an STT-LLM-TTS pipeline, reliable turn detection, handling interruptions, and LLM orchestration. Notable pieces: