LiveKit Agents Guide 2026: Build Voice AI You Control

LiveKit Agents is an open-source framework for building voice AI agents in Python or Node.js: your code joins a realtime LiveKit room as a participant and streams the caller's audio through speech-to-text, a language model and text-to-speech, or a realtime speech model. You can run it on LiveKit Cloud, which charges per agent session minute, or host the whole stack yourself under the Apache 2.0 license.

Facts and prices checked October 2026 against the LiveKit Agents documentation, LiveKit Cloud pricing, LiveKit billing docs, LiveKit telephony docs and Vapi pricing.

Hosted platforms such as Vapi, Retell and Bland get you a working phone agent fastest. LiveKit asks for more engineering in exchange for two things: a much smaller platform fee per minute and control over where audio runs and which models you use. This guide explains how the framework fits together, what LiveKit Cloud costs, and, with clearly labelled illustrative math, where the crossover sits against a hosted platform and against self-hosting. For the wider map of phone-agent options, start with our voice AI hub on building AI phone agents.

Who LiveKit Agents is for

• Developers building a voice product that will run thousands of minutes a month and needs custom logic, tools and integrations in code.
• Teams with data rules that want audio and transcripts on their own infrastructure, or a choice of model providers per region.
• Agencies standardising on one stack for many client agents, where a per-minute markup across every client adds up.

It is not the best first step if you need a receptionist live this afternoon and nobody on the team writes Python or TypeScript. In that case, start with a hosted platform; our Bland AI vs Vapi vs Retell comparison covers the trade-offs, and you can move to LiveKit once volume justifies the work.

How LiveKit voice AI works end to end

LiveKit's documentation describes the agent as a stateful realtime bridge between AI models and users. Your agent code registers with a LiveKit server (Cloud or self-hosted) as an agent server. When a room is created, the server dispatches a job, a subprocess that joins the room. Users connect over WebRTC from a web or mobile app, or over the phone through LiveKit's SIP service. The agent talks to your backend over HTTP and WebSockets.

Inside the job, an AgentSession orchestrates the conversation. The framework handles the hard realtime parts: streaming audio through the pipeline, deciding when the caller has finished speaking with its turn detection model, stopping playback when the caller interrupts, and calling tools you define. Multi-agent handoffs let you split a flow into stages, for example a triage agent that hands off to a booking agent with the context preserved. Plugins cover most major model providers, and LiveKit Inference on Cloud lets you call models without managing separate API keys.

Build your first LiveKit agent: the steps

The official Voice AI quickstart gets a basic assistant running in Python or Node.js. The shape of the work, from blank folder to a phone number that answers, looks like this:

Keep the first version narrow: one job, such as answering opening-hours questions and booking a slot, with a transfer to a human for everything else. Turn detection and latency matter more to callers than clever prompts, so test with real phone audio early. The STT and TTS choice also changes how the agent sounds and how quickly it replies.

Tools are where most of the value sits. A tool is a function your agent can call mid-conversation; the patterns are the same as LLM function calling elsewhere, which our guide to Claude API function calling walks through.

Phone calls: SIP, trunks and dispatch rules

LiveKit's telephony docs add two objects to rooms and participants. Trunks connect your SIP provider to LiveKit: inbound trunks accept calls and can be restricted by IP or number, and outbound trunks place calls. Dispatch rules decide which room each inbound caller lands in. Outbound calls are made by creating a SIP participant through the API. Supported features include DTMF, cold transfer via SIP REFER, warm transfer and caller ID. Connectors also bridge Twilio calls and WhatsApp Business calls into rooms without a SIP trunk.

If you place outbound sales calls with an AI voice, consent rules apply in the US: the FCC has confirmed that AI-generated voices count as artificial voices under the Telephone Consumer Protection Act. The outbound section of our platform comparison links the official FCC and FTC pages.

LiveKit Cloud pricing

From LiveKit's pricing page, checked October 2026. The billing docs add that agent session minutes are metered in one-minute increments from when the agent joins the room, and that on the free Build plan the allowance is a hard cap rather than an overage.

Model costs sit on top, either through LiveKit Inference or your own provider keys. The pricing page lists per-minute equivalents for Inference models; for example, Deepgram Nova-3 at $0.0048 a minute for speech-to-text on Build and Ship, GPT-5.4 mini at $0.0030 a minute and Inworld Realtime TTS 2.0 Flash at $0.0090 a minute. Carrier charges from your SIP provider are separate too. Session recording and observability have their own allowances per plan.

LiveKit vs Vapi: platform fee and control

Vapi's pricing page lists $0.05 per minute for hosting, with models passed through at provider cost and telephony billed by the carrier (checked October 2026). LiveKit Cloud's comparable platform charges are $0.01 per agent session minute plus $0.003 to $0.004 per third-party SIP minute after the plan allowance. Both let you pick models, so the model bill can be held equal; the difference is the platform layer and how much you build yourself.

The cost crossover: illustrative math

This section is an illustrative calculation from published rates, not a measurement. It compares only the platform layer, because model and carrier costs are similar whichever platform you choose. The self-hosted figure is an assumed input for servers, monitoring and the engineering time to run them; replace it with your own.

What the table says, on these inputs. LiveKit Cloud's platform layer is cheaper than Vapi's hosting fee from low volumes, but at 10,000 minutes the $380 difference will not pay for much engineering, so convenience can rightly win. By 100,000 minutes a month the gap is several thousand dollars and the build effort starts to pay. Against an assumed $4,000 self-hosting budget, Vapi crosses over at 80,000 minutes (80,000 x $0.05), while LiveKit Scale only crosses at about 319,000 minutes ($500 plus 269,000 x $0.013). Ship's 20 concurrent sessions would also cap a busy line well before the higher volumes, which pushes you to Scale anyway.

Two caveats. First, the self-hosted line is the weakest number here: one engineer's time on call can exceed it, and a second region doubles parts of it. Second, data control is not a cost line. If a contract requires audio to stay inside your network, self-hosting can be the answer at any volume. For the open-source side of the stack in general, our best voice AI platforms roundup covers how LiveKit compares with the other developer options.

What self-hosting actually involves

LiveKit's pricing FAQ states that the Agents framework and the media server are open source and can run on your own infrastructure. The self-hosted SIP server docs show the moving parts: a LiveKit server, a SIP server and Redis, with the SIP signalling port 5060 and media ports 10000 to 20000 reachable from the internet. On top of that you run the agent servers, scale them for long-lived sessions, drain them during deploys and keep logs and transcripts somewhere.

A common middle path is to build on LiveKit Cloud from day one, since the same agent code runs on either, and move to self-hosting only once the numbers and the team are there. If you already self-host other automation tools, our guide to running n8n with Docker Compose covers similar container and networking basics.

Mistakes that cost the most time

• Testing only in the browser. Phone audio is narrower and noisier than WebRTC. Test over a real trunk before judging the STT and turn detection.
• Leaving agents connected. Billing runs until the room ends or the agent disconnects; LiveKit's billing docs recommend calling ctx.shutdown() to end a session explicitly.
• Choosing the biggest model by default. A larger model adds latency and cost per minute. Start small, then upgrade the parts of the flow that need it.
• No human fallback. Every production line needs a transfer path and a plan for when the agent cannot help.

If you are turning voice agents into a product for clients, our AI SaaS Builder program covers scoping, pricing and shipping automation products.

LiveKit Agents: FAQ