CustomVapi

Basics

Name System prompt who the assistant is and how it behaves; {{variables}} allowed

Built-in variables: {{now}}, {{date}}, {{time}}, {{day_of_week}}, {{year}}. Others come from the widget or SDK.

Engine

How the assistant hears and speaks Realtime speech-to-speech (fastest, most natural) Cascade: transcribe → ChatGPT → speech (cheapest, most control)

Voice

Conversation

Caller can interrupt the assistant Smart turn detection (waits for the caller to finish their thought) Wait a little longer after very short phrases ("Hi there.") Ignore background voices much quieter than the caller (TV, people nearby)

Call behaviour

Assistant can hang up when the conversation is done Apology when a provider fails

Knowledge & tools

Knowledge bases documents the assistant answers from Use knowledge by Automatic: in the prompt when small (up to ~16k characters), else search Prompt: fastest answers (up to ~48k characters) Search: any size, each lookup adds 1–3 s Tools JSON array; calls are POSTed to webhook_url, or run in the caller's page with "type": "client"

Post-call analysis

Summarize and evaluate every call Summary instructions Success question Data to extract JSON schema, e.g. name, phone, appointment time

Webhooks

Server URL receives call events Events Signing secret adds X-CustomVapi-Signature (HMAC-SHA256)

Privacy & security

Record calls (tell callers they're being recorded) Require a call token from your backend (POST /api/call-tokens; its variables can't be edited by callers) Websites allowed to start calls one per line, e.g. https://www.example.com; empty = any Max simultaneous calls

Embed on a website