Skip to content
agentscamp
Tool

Livekit

Open-source realtime infrastructure — a WebRTC server plus the LiveKit Agents framework for production voice AI, with turn detection, telephony, and cloud.

freemiumvoice
Updated Jun 11, 2026
webrtcvoice-agentsrealtimeopen-sourcetelephony
Visit website Source
View as Markdown

LiveKit is the open-source realtime stack voice AI standardized on: an Apache-2.0 WebRTC server plus the LiveKit Agents framework (Python/Node) wiring STT→LLM→TTS or speech-to-speech models, with an open multilingual turn-detection model, full telephony (SIP, DTMF, transfers), and LiveKit Cloud as the managed network. Self-host free; cloud freemium with metered minutes.

LiveKit became voice AI's load-bearing infrastructure the unglamorous way: by being the open-source WebRTC stack that worked, then building the agent layer the moment voice agents needed one. The credential says it all — ChatGPT's Voice Mode runs over LiveKit, per LiveKit's own announcements.

Highlights

  • The WebRTC core — an Apache-2.0 SFU for low-latency audio/video at scale; self-host anywhere, in Go.
  • LiveKit Agents — the framework (Python and Node) for production voice agents: pluggable STT/LLM/TTS pipelines or realtime speech-to-speech models, with interruption handling built in.
  • Open turn detection — a multilingual semantic turn-detection model (13 languages, ~25ms CPU inference) — the hardest part of natural conversation, open-sourced.
  • Telephony 1.0 — SIP, DTMF, transfers, thousands of concurrent calls: the phone-system half most stacks bolt on late.
  • LiveKit Inference — one interface routing STT/LLM/TTS across providers (Cartesia, Deepgram, and friends plug in).
  • Cloud when you want it — serverless agent deployment, observability (session replays, traces), a real free tier, metered minutes beyond.

In an AI-assisted workflow

pip install "livekit-agents[openai,silero,deepgram,cartesia,turn-detector]"
# define an Agent with your STT/LLM/TTS mix (or a realtime model) and deploy —
# self-hosted server or LiveKit Cloud

It's the substrate under the voice-agent pipeline: you bring the models, LiveKit owns transport, turns, and telephony — the parts that make demos fall over in production.

NOTE

Use livekit.com (the .io domain redirects), and pin agent-framework versions — the 1.0 redesign retired pre-1.0 patterns and the cadence stays brisk. Self-hosting is genuinely free but re-creates what Cloud meters (TURN, scaling, orchestration).

Good to know

Apache-2.0 throughout (~19k/11k stars across server/agents), with a $45M Series B (April 2025) and a $100M Series C at a $1B valuation (Index Ventures, January 2026) — agents downloads topped a million a month. The build-vs-buy line against Vapi and the OSS-pipeline comparison with Pipecat are drawn in Realtime Voice Agents.

Frequently asked questions

What exactly is LiveKit — the server or the agents framework?
Both, layered: livekit/livekit is the WebRTC SFU (the media transport), and livekit/agents is the framework for building voice agents on top — pluggable STT/LLM/TTS providers or realtime speech-to-speech models, turn detection, interruptions, telephony. Self-host both for free, or run on LiveKit Cloud and pay metered minutes.
Does ChatGPT's voice mode really run on LiveKit?
Per LiveKit's own blog: yes — OpenAI integrates a LiveKit client SDK in the ChatGPT app, with calls connecting over LiveKit Cloud, alongside an announced partnership around the Realtime API. It's vendor-attributed (phrase it that way), but it's the strongest production credential in the category.
LiveKit vs Vapi — build or buy?
LiveKit is infrastructure you assemble: maximum control, open source, your provider choices, real engineering investment. Vapi is the managed assembly: an API where agents exist in minutes at a platform fee per minute. Teams with realtime engineering appetite (or scale economics) build on LiveKit; teams shipping voice features fast buy Vapi.

Related