Build

Agents

Agents are the AI voice and chat workers you build in Callable — they answer and make phone calls, chat on your website, and are grounded in your business info, knowledge base, and tools.

What an agent is

An agent bundles everything the AI needs to hold a conversation: an identity and system prompt, a voice, a transcriber, grounding content, callable tools, and an optional conversational flow. Once built, you attach it to a phone number (to answer inbound calls or dial outbound campaigns) or embed it as a website chat widget.

Under the hood, every conversation runs the same loop: the caller (or visitor) speaks or types, the input is transcribed, the agent's model generates a grounded reply, and — for voice — that reply is streamed back as speech sentence-by-sentence so the person can barge in and talk over the agent.

Creating an agent

  1. In the left nav, open Agents.
  2. Start a new agent.
  3. Choose the agent type — Voice or Chat (see below).
  4. Fill in the Agent tab: your business name, industry, hours, services, FAQs, and the greeting the agent opens with. You can also edit the full system prompt.
  5. Configure the remaining tabs as needed (Voice, Transcriber, Knowledge Base, Tools, Flow, Model, Analysis).
  6. Save, then test.
You choose the agent type once, at creation. Pick based on where the agent will live — on the phone, or on your website.

Voice vs chat

Callable has two agent types, chosen when you create the agent:

  • Voice agents answer and make phone calls. They use a transcriber (speech-to-text) and a voice (text-to-speech) on top of the model, and they attach to a phone number.
  • Chat agents talk over text on your website. Every chat agent gets a public chat page and can be embedded as a widget. They run the same model pipeline as voice agents, minus the speech steps.

The builder shows the tabs relevant to the type you picked — for example, the Voice and Transcriber tabs matter for voice agents.

A tour of the builder tabs

The agent builder is organized into tabs. Here's what each one does; follow the links for full details.

Agent

Your agent's identity and grounding basics. Set the business name, industry, hours, services, and FAQs, write the greeting (the opening line the agent says or sends), and edit the fully customizable system prompt that steers its behavior. See Agent tab.

Voice

For voice agents: choose the text-to-speech provider — Cartesia (the platform default, with Sonic voices) or ElevenLabs (bring your own key) — pick a voice from the library, and set speed (0.5–2.0), language, and fallback voices used if the primary voice fails. See Voice.

Transcriber

For voice agents: the speech-to-text settings, powered by Deepgram. Tune the model (nova-2-phonecall is the phone default), endpointing (trailing silence before the agent decides you've finished), keyterm boosting, background denoising, confidence threshold, and language. See Transcriber.

Knowledge Base

Ground the agent in your own content. Add website URLs (fetched), upload files (PDF/docs), or paste text. The agent draws on this content on the widget, on the phone, and in the dashboard. See Knowledge Base.

Tools

Give the agent abilities it can call mid-conversation. Attach tools — including built-in transfer-to-human, send SMS, and end-call — and add MCP servers as callable tools. See Tools.

Flow

Script structured conversations with the Conversational Flow node-graph editor: define nodes, transitions, and variables, use Transfer and End nodes, and block interruptions per node. The flow advances one step per turn — ideal for something like an appointment-booking sequence. See Conversational Flow.

Model

Choose the LLM that powers the agent. A per-agent model override takes precedence; otherwise the owner's default is used. Provider dropdowns only show providers that are actually connected. See Model.

Analysis

Configure the post-conversation analysis pass — how the agent's calls and chats are summarized and how fields are extracted afterward. See Analysis.

Saving and testing

  • Save your changes in the builder before testing — saving persists the configuration across all the tabs.
  • Test the agent right away: use the built-in test to try a chat, or place a call to confirm the greeting, voice, tools, and flow behave as expected.
Tip: if you change tools, voice, or model settings on an existing agent, re-save the agent so the live conversation picks up the new configuration.

Next steps