Phone Agent

A Sunrise Travel front-desk receptionist running on Pipecat's PhoneLLM Mini, a model tuned for voice agents: sub-100 ms time to first token, reliable tool calls, and no reasoning pass to wait through. Ask about the weather where you're headed and it looks it up live mid-conversation.

Voice

Talk instead of type

Browser voice: your browser transcribes and speaks; PhoneLLM Mini writes the replies

Last turn latency

Measured in the browser from send, through the proxy.

First event: —
Tool call: —
First word: —
Total turn: —

Realtime voice (optional)

Runtype's server-side speech pipeline needs a client token and a saved agent set to phonellm-mini. Leave blank to use the browser's speech recognition and voices; typing always works without them.

Notes

The model

PhoneLLM Mini Alpha 2 from Pipecat (Daily), served on Cloudflare Workers AI and routed by Runtype as phonellm-mini. It is post-trained from Nemotron 3 Nano 30B-A3B for phone and voice agents: fast tool calling, short conversational replies, no reasoning tokens. Runtype forces temperature to 0 and disables thinking server-side. Pricing is $0.06 / $0.24 per million input / output tokens.

How it works

The widget posts to /api/chat/dispatch-phone-agent on the demo proxy, which pins the agent server-side (PHONE_RECEPTIONIST_AGENT in @runtypelabs/persona-proxy) and keeps the Runtype API key off the page. PhoneLLM Mini is a text model: it never hears or produces audio itself. With voice on, the browser's Web Speech API transcribes you and its built-in voices read each reply aloud; supply a client token and agent id to switch speech-to-text and text-to-speech to Runtype's realtime voice pipeline instead.

The weather tool

A keyless external GET against Open-Meteo. The model fills in coordinates from the place name itself.

{
  name: "get_current_weather",
  toolType: "external",
  parametersSchema: {
    type: "object",
    properties: {
      latitude: { type: "number" },
      longitude: { type: "number" },
    },
    required: ["latitude", "longitude"],
  },
  config: {
    method: "GET",
    url: "https://api.open-meteo.com/v1/forecast?latitude={{latitude}}&longitude={{longitude}}&current=temperature_2m,weather_code&temperature_unit=fahrenheit",
  },
}

A second tool, check_office_hours, reads the office's local time from timeapi.io so the receptionist can say whether it's open.

What to look for

A one-line holding phrase ("Let me check the weather in Austin for you"), then the tool bubble, then the answer: the model looks it up rather than guessing, and fills the silence the way a receptionist would.
Replies are one or two short sentences with numbers spelled out, written to be spoken.
No reasoning bubble: the model answers without a thinking pass.
The latency readout: the first words arrive before the tool call even starts, and the whole turn, live weather lookup included, finishes in a few seconds.

Caveats

PhoneLLM Mini is an alpha model and speaks English only. Expect occasional rough edges on long or multi-part requests; it is built for short phone-style exchanges.