LiveKit

Our full take

LiveKit is the layer most people never see. When a voice agent and a human talk, something has to carry the audio between them with low enough delay that the call feels like a call, and route it reliably when thousands run at once. That is what LiveKit does. It is the real-time transport, plus a framework that lets you wire your own speech-to-text, language model and voice on top. OpenAI uses it to move the audio for ChatGPT’s Advanced Voice between your phone and its servers, which tells you the scale it runs at.

So the first thing to be clear about: this is not a finished agent you switch on. It is the plumbing and the toolkit. You assemble the agent yourself. For a developer that is the appeal, you own every piece and swap any of them. For a non-technical buyer who wants to plug in and book appointments by Friday, this is the wrong shelf. Start with Vapi or Synthflow instead, and come back to LiveKit when you have outgrown them.

Here is what you actually get. The open-source core is two things: the LiveKit server (the media router, written in Go) and the Agents framework (Python and Node.js). Both are Apache 2.0 licensed, so self-hosting them costs you nothing beyond the servers you run them on. The framework handles the awkward real-time problems for you: streaming audio through a speech-to-text then language-model then voice pipeline, working out when the caller has stopped talking (turn detection), handling interruptions when they talk over the agent, and built-in telephony so a person can join from an ordinary phone line rather than a web app. That telephony piece matters. Plenty of “building-block” vendors leave you to bolt on Twilio yourself. LiveKit ships inbound and outbound phone support, over SIP, in the box.

What it deliberately does not give you is the speech, the brain or the voice. You bring those. Deepgram or OpenAI for the listening, OpenAI or Anthropic for the thinking, ElevenLabs or Cartesia for the talking, all plug in as provider plugins. That is the flexibility and the homework in one sentence. Nobody locks you to a house voice, and nobody hands you a working stack either.

Now the money, because LiveKit prices in a way that trips people up. Self-hosting the open source is free. The managed service, LiveKit Cloud, is where the per-minute numbers live, and they are not one number. Cloud meters the agent session (the orchestration running your call) at about $0.01 a minute once you are past the included allowance. On top sit two smaller meters: the WebRTC participant minutes that carry each person on the call (roughly $0.0004 to $0.0005 a minute, genuinely tiny) and bandwidth at $0.10 to $0.12 a gigabyte (a voice call barely touches this). If you route real phone calls, LiveKit’s own US local inbound runs about $0.01 a minute, third-party SIP is cheaper at $0.003 to $0.004, toll-free is $0.02, and phone numbers are $1 a month ($2 for toll-free).

Add it up and LiveKit’s own meters land around $0.02 a minute for a phone-connected agent. That is the cheap part, and it is misleading on its own, because it does not include the bits that cost the most. Your speech-to-text, language model and voice providers are billed by them, not by LiveKit, usually another $0.06 to $0.15 a minute depending on which models you pick. So the honest all-in for a Cloud voice agent sits around $0.08 to $0.20 a minute, most of which is the providers you chose, not LiveKit’s cut. If you would rather not sign up with each provider separately, LiveKit will resell managed inference through its own credits, which simplifies the bill at the cost of a small markup.

The plans themselves are Build (free, no card, 1,000 agent minutes), Ship ($50 a month, 5,000 agent minutes), Scale ($500 a month, 50,000 agent minutes) and Enterprise (custom). The tiers mostly buy you more included minutes and higher concurrency, from 5 simultaneous agents on Build up to around 600 on Scale.

On compliance LiveKit is in good shape, with one catch on where it starts. Its trust page states an audited SOC 2 Type II, GDPR with a data processing addendum on every tier, and EU-US Data Privacy Framework certification. A signed HIPAA Business Associate Agreement and region pinning are real, but they begin on the Scale tier ($500 a month), so a healthcare build cannot use the free plan and stay compliant. ISO 27001 and PCI DSS are listed as in progress, not done, so do not count on them yet.

The 1 to 10 scores here are an editorial preview, our provisional read to get the framework in place, not a measured result. We have not run LiveKit through our own call tests yet, so there is no Voxrater latency figure on this page. The pricing, licence, telephony and compliance detail is sourced from LiveKit’s pricing page, Agents docs, GitHub repository and trust page, captured 2026-05-31.

How this page was made. AI-assisted research, sourced against LiveKit’s own pricing, documentation and trust pages, with editorial judgement and review by Voxrater. Last reviewed 2026-05-31. We have not yet call-tested LiveKit, so no benchmark latency is shown. Spot an error? Every figure links to a dated source above. Tell us and we will correct it.