Quickstart

By the end of this page you will have a voice agent you can talk to in your browser. It needs nothing installed and no credit card - a new organization starts with a welcome credit.

Create an agent

In the dashboard, go to Agents and create one. Choose Prompt for a single-purpose agent (this guide) or Workflow when the conversation has branches, tools and handoffs - see Workflows.

Write the prompt

The prompt is the agent's instructions, and it is where most behaviour lives. Something like:

You are the receptionist for Acme Dental. Greet the caller, find out whether
they want to book, move or cancel an appointment, and collect their full name
and phone number. Keep replies short and natural. If they ask about anything
else, say a human will call back.

Write it as if briefing a new employee: what they do, what to ask, what to refuse.

Choose who speaks first

Set first speaker to assistant if the agent should greet, or user if it should wait. This is the single most common cause of an agent that "does nothing" or talks over the caller at the start.

Test it

Open the Playground, pick your agent and start a call. Talk to it. Below the conversation you can see which step it is on and what it captured.

If something is off, edit the prompt and start a new session. You can also try a change without saving it - see the Playground.

Ship it

The same agent runs anywhere without a rewrite:

Defaults you can ignore for now

An agent works out of the box. The model, transcriber and voice all fall back to platform defaults, and you can change them later on the agent's Models and voices settings - see Models and voices.

Leave the transcriber language on multi (auto-detect) unless you are certain every caller speaks one language. A fixed, wrong language is the most common cause of misheard names and numbers.

  • Core concepts - the vocabulary the rest of the docs uses.
  • Prompt agents - every setting a prompt agent has.
  • Tools - let the agent look things up and take actions.
Quickstart · Glytos Docs