LiveKit

A voice agent is four vendors pretending to be one product: something to carry audio in real time, something to turn speech into text, a model to think, and something to turn text back into speech. Add a phone number and it’s five. LiveKit is the piece in the middle — open-source WebRTC transport plus an agents framework — and LiveKit Cloud sells the other four through one bill, which is why it’s the pick here.

LiveKit Cloud link

Overage is $0.01/min for agent sessions and US inbound calls on every plan; SIP trunking from a third-party carrier is $0.004/min on Ship and $0.003/min on Scale.

That BAA column is the one to read twice. Scale at $500 is not HIPAA-eligible; if you handle PHI, budget for an Enterprise conversation from day one, or self-host the open-source server and take the compliance on yourself.

LiveKit Inference link

Per-minute prices for the models LiveKit resells, so a whole turn is one line item:

Rough total for a cheap stack: about $0.02–0.05 per agent-minute all in, dominated by TTS. The LLM is a rounding error at these rates; the voice is not.

Going direct instead link

Inference is a convenience markup. The same vendors sell to you directly, and going direct is what you’ll do the moment you need a BAA, since LiveKit’s only covers LiveKit.

Two patterns fall out of that table. For a hobby or startup product, LiveKit Cloud Build/Ship plus Inference is the shortest path and the cheapest bill. For anything regulated, the stack rearranges around who will sign: Azure for the model and speech, Deepgram or Azure for STT, Twilio for the phone line, your own Postgres for data, and LiveKit Enterprise or self-hosted in the middle.

Prices are as of September 2026 and every one of these vendors reprices quarterly — check before you put a number in a proposal.

First agent link

  pip install "livekit-agents[openai,deepgram,cartesia,silero]"
export LIVEKIT_URL=wss://<project>.livekit.cloud
export LIVEKIT_API_KEY=... LIVEKIT_API_SECRET=...
  

  from livekit.agents import Agent, AgentSession, JobContext, WorkerOptions, cli
from livekit.plugins import deepgram, openai, cartesia, silero

async def entrypoint(ctx: JobContext):
    session = AgentSession(
        stt=deepgram.STT(model="nova-3"),
        llm=openai.LLM(model="gpt-4o-mini"),
        tts=cartesia.TTS(),
        vad=silero.VAD.load(),
    )
    await session.start(agent=Agent(instructions="Be brief."), room=ctx.room)

if __name__ == "__main__":
    cli.run_app(WorkerOptions(entrypoint_fnc=entrypoint))
  

  python agent.py dev          # then talk to it in the LiveKit Agents Playground
  

Swapping a vendor is changing one constructor. That is the practical reason to build on LiveKit even if you never use Inference: the compliance-driven rearrangement above is a four-line diff, not a rewrite.

Next link

The infrastructure works. Now get paid → Lemon Squeezy