LiveKit: voice agents that answer in 300ms. — DeltaForceOS

◢ The stack

LiveKit Cloud

Realtime voice infra for sub-300ms agent calls

Free 10k min · pay-as-you-go

Deepgram

Realtime speech-to-text, lowest-latency transcription

$200 free credit

ElevenLabs

Text-to-speech for your agent's voice

Free 10k chars · $5/mo

Anthropic API

Claude API for the smartest agent reasoning

$5 free credit · pay-as-you-go

◢ The build · 4 steps · 18 min

Follow these in order. Don't skip.

Step 01 / 04

Create your LiveKit Cloud project

Create a LiveKit Cloud project — livekit.io → Sign up → Create Project. Pick the region closest to your users.

LiveKit API keys page — Project → Settings → Keys. Create a new key. Copy the URL, key, secret.

.env.local

1LIVEKIT_URL=wss://your-project.livekit.cloud
2LIVEKIT_API_KEY=APIxxxxxx
3LIVEKIT_API_SECRET=secretxxxxxx
4DEEPGRAM_API_KEY=...
5ELEVENLABS_API_KEY=...
6ANTHROPIC_API_KEY=...

Step 02 / 04

Install the LiveKit Agents Python SDK

The Agents SDK handles the audio loop — capture, send to STT, get LLM response, stream to TTS — so you write one Python file.

Terminal

1# In a fresh agent/ folder
2python3 -m venv .venv
3source .venv/bin/activate
4 
5pip install livekit-agents \
6  "livekit-plugins-anthropic" \
7  "livekit-plugins-deepgram" \
8  "livekit-plugins-elevenlabs" \
9  "livekit-plugins-silero" \
10  python-dotenv

Step 03 / 04

Write the agent in ~40 lines

agent/main.py

1import asyncio
2from dotenv import load_dotenv
3from livekit import agents
4from livekit.agents import AutoSubscribe, JobContext, llm
5from livekit.agents.voice_assistant import VoiceAssistant
6from livekit.plugins import anthropic, deepgram, elevenlabs, silero
7 
8load_dotenv()
9 
10async def entrypoint(ctx: JobContext):
11    initial_ctx = llm.ChatContext().append(
12        role="system",
13        text=(
14            "You are a friendly receptionist. Answer in 1-2 short sentences. "
15            "If the caller wants to book, ask for their name and preferred day."
16        ),
17    )
18 
19    await ctx.connect(auto_subscribe=AutoSubscribe.AUDIO_ONLY)
20 
21    assistant = VoiceAssistant(
22        vad=silero.VAD.load(),
23        stt=deepgram.STT(model="nova-3"),
24        llm=anthropic.LLM(model="claude-sonnet-4-6-20250101"),
25        tts=elevenlabs.TTS(voice="Rachel"),
26        chat_ctx=initial_ctx,
27    )
28 
29    assistant.start(ctx.room)
30    await asyncio.sleep(1)
31    await assistant.say("Hi — thanks for calling. How can I help?", allow_interruptions=True)
32 
33 
34if __name__ == "__main__":
35    agents.cli.run_app(agents.WorkerOptions(entrypoint_fnc=entrypoint))

Step 04 / 04

Run the agent locally and talk to it

Terminal

1# Run the agent worker (it connects to LiveKit Cloud and waits for rooms)
2python agent/main.py dev

▸Now go to playground.livekit.io
▸Sign in with the same project's URL + key
▸Click "Connect" → it joins a room and your agent picks up
▸Speak into your mic — the agent responds in 300–600ms

◆ You're done when

That's it. Your voice agent works. Hook it to Twilio for phone numbers, plug in tools (calendar, CRM) via function calling, deploy it as a worker on Render or Fly.

◆ Ship-it checklist

5 CHECKS

LiveKit Cloud project created, keys in .env.local
Deepgram + ElevenLabs + Anthropic keys set
Python venv with livekit-agents installed
agent/main.py runs without errors
You spoke to the agent through playground.livekit.io and it responded

← All guides Show your build in the community →