the classic STT → LLM → TTS pipeline. One model listens and speaks directly — lower latency, more natural prosody, and native barge-in handling.

Enable it#

Set the server-side LLM to a realtime model. That's it:

const agent = pc.agent("receptionist", {
  llm: "pinecall/gpt-realtime",
  instructions: "You are a warm, concise phone receptionist for Bella Vista.",
  greeting: "Hi! Thanks for calling Bella Vista — how can I help?",
  tools: [checkReservation],
});
agent.addChannel("phone", "+19035551234");

Any model whose name contains realtime activates the mode (gpt-realtime, gpt-realtime-2.1, gpt-realtime-2.1-mini, …).

What changes under the hood#

	Classic pipeline	Realtime S2S
STT	Deepgram / Gladia / …	built into the model
Turn detection / VAD	smart_turn / silero	built into the model
TTS	ElevenLabs / Cartesia / …	built into the model
Barge-in	server energy + VAD	native (model interrupts itself)
Billing	STT min + LLM tokens + TTS chars	LLM audio tokens only (+ telephony)

Because the model owns the whole audio path, stt, tts/voice, turnDetection, vad and interruption config are ignored in this mode. The speaking voice is one of OpenAI's realtime voices — default marin, override with:

config: { realtime: { voice: "cedar" } }

What still works exactly the same#

Channels: phone (Twilio) and WebRTC (browser) — same addChannel calls.
Greeting: the greeting: field (or one call.say on call.started) — the realtime model speaks it verbatim.
Tools: same tools definitions, same SDK execution — the server sends llm.tool_call, your handlers run locally, results go back mid-call.
Events: call.started, user.message, turn.end, bot.speaking, bot.finished, bot.interrupted, call.ended (with the full transcript).
call.say(text): speaks the exact text through the realtime voice.
Prompt vars: built-in {{date}}, {{time}}, {{day}}, {{datetime}} resolve at session start (with the agent's timezone).

Not supported (yet) in realtime mode#

Skills (skill() progressive disclosure)
Knowledge bases / RAG (knowledgeBase)
Per-word captions (bot.word) and hold music
Custom setPromptVars refresh per turn (vars resolve once at call start)

If you need any of these, stay on the classic pipeline for that agent.