Two SDKs.
Same shape as OpenAI Realtime.

# npm
npm install @toolkit-llm/voice

# or pnpm / yarn / bun
pnpm add @toolkit-llm/voice
yarn add @toolkit-llm/voice
bun add @toolkit-llm/voice

import { Toolkit } from "@toolkit-llm/voice";

const client = new Toolkit({
  apiKey: process.env.TOOLKIT_API_KEY,
});

const session = await client.realtime.sessions.create({
  model: "toolkit-voice",
  instructions: "You are a helpful auto-dealer agent named Tony.",
  voice: "VARM1",
  modalities: ["text", "audio"],
});

session.on("response.audio.delta", (event) => {
  // event.delta is base64 PCM16 (or opus if you set audio_format)
  speakers.write(event.delta);
});

session.on("conversation.item.input_audio_transcription.completed", (event) => {
  console.log("Caller said:", event.transcript);
});

session.appendAudio(micChunk);

const call = await client.calls.create({
  to: "+14155551234",
  from: "+14152278669",  // Your provisioned number
  persona: "tony",
  max_seconds: 60,
});

// Stream events as the call progresses
for await (const event of call.events()) {
  if (event.type === "call.ended") break;
}

import { verifyWebhook } from "@toolkit-llm/voice/webhooks";

// Express
app.post("/voice/event", express.raw({ type: "*/*" }), (req, res) => {
  const event = verifyWebhook({
    body: req.body,                                  // raw bytes
    signature: req.header("X-Toolkit-Signature"),
    secret: process.env.TOOLKIT_WEBHOOK_SECRET,
    tolerance_s: 300,                                // 5min replay window
  });
  // event is now a typed Event union
});

pip install toolkit-voice

# Or with uv (faster):
uv pip install toolkit-voice

import asyncio
from toolkit import Toolkit

async def main():
    client = Toolkit()
    session = await client.realtime.sessions.create(
        model="toolkit-voice",
        instructions="You are a helpful auto-dealer agent named Tony.",
        voice="VARM1",
    )

    @session.on("response.audio.delta")
    def on_audio(event):
        # event.delta is base64 PCM16
        speakers.write(event.delta)

    @session.on("conversation.item.input_audio_transcription.completed")
    def on_transcript(event):
        print(f"Caller said: {event.transcript}")

    await session.append_audio(mic_chunk)

asyncio.run(main())

call = client.calls.create(
    to="+14155551234",
    from_="+14152278669",
    persona="tony",
    max_seconds=60,
)

async for event in call.events():
    if event.type == "call.ended":
        break

from fastapi import FastAPI, Request, HTTPException
from toolkit.webhooks import verify_webhook, InvalidSignature

app = FastAPI()

@app.post("/voice/event")
async def voice_event(request: Request):
    raw_body = await request.body()
    try:
        event = verify_webhook(
            body=raw_body,
            signature=request.headers["X-Toolkit-Signature"],
            secret=os.environ["TOOLKIT_WEBHOOK_SECRET"],
        )
    except InvalidSignature:
        raise HTTPException(401, "Invalid signature")
    return {"ok": True}