AI voice agents
AI voice agents that answer and make calls
One agent handles your inbound calls and makes outbound calls, in English and Hindi today and in Arabic where the voices are wired. It books appointments, qualifies leads, answers support questions and hands the call to a person when it should. Run it hosted, with your own keys, or on your own server.
No fake numbers on this page. Where a feature is new or still being finished, we say so.
Live demo
Try our live demo
Receive a live call from our agent and discover how an AI caller answers, asks questions, and books. Pick a use case, leave your number, and your phone rings — no sign-up, no sales rep.
Design your agent's workflow visually — anyone can build this, no code.
Every demo on this page runs on a flow like the one above. Drag, drop, publish — this is the exact builder waiting inside your Callingtap workspace.
Demo agent types
Choose a demo agent type.
What the agents do
The work an agent does on a call, start to finish
Receptionist
Answers the phone, checks calendar availability, books the appointment and reads the confirmation back, transfers to a person, or takes a message. Built from the receptionist templates, which include the after-hours, multi-department and law-firm versions.
Appointment booking and rescheduling
Books, confirms and moves appointments, with reminders that cut no-shows. The rider booking, IVR scheduling and reminder templates do this today.
Lead qualification
Asks the qualifying questions, scores intent, and books a demo or hands a hot lead to a person. Four lead templates cover basic, high-intent, B2B demo and screener calls.
Customer support
Answers from your knowledge base, with the source cited, and looks a Shopify order or a Calendly booking up during the call. If the document does not have the answer, the agent says so.
Abandoned cart recovery
Calls a customer who left items in a cart, confirms the items, and helps them finish the order or answers a question - without pressing them.
KYC and onboarding
Walks a customer through identity checks and onboarding steps, and never asks for a full password, PIN or card number.
Inbound and outbound
In on your number, out on your campaigns
Inbound
Connect a phone number through Twilio, Telnyx, Plivo, Vonage or SIP, and the agent answers as your receptionist. Business hours decide what happens after hours: take a message, transfer, or keep answering.
Outbound
Dialer campaigns call your list in Standard mode, or auto-enrol new contacts by tag. Campaigns run on the dialer queue; a contact that is not on the campaign cannot be dialled by a workflow - the page says so rather than pretending otherwise.
One agent, every channel
The same agent, not a copy of it
The agent is one definition - one prompt, one set of tools, one knowledge base. It runs on voice, WhatsApp, web chat, the website widget, SMS and email. A conversation that starts on WhatsApp and moves to a call keeps its context.
Voice
Inbound and outbound calls over Twilio, Telnyx, Plivo, Vonage or SIP, with the same prompt, guardrails and playbook.
Official Meta Cloud API, through the same agent as voice. Inside the 24-hour window the agent replies as itself; outside it, only an approved Meta template. One agent per workspace on WhatsApp.
Chat and web
The website widget and webchat run the same agent, so answers stay consistent with what a caller hears.
SMS and email
The same agent answers SMS and email threads, with the same knowledge and guardrails.
Shared inbox
Every conversation is in one place. Your team can see the thread, add a note, assign it, or take it over with a warm handoff - without losing context.
Warm handoff
When the agent cannot resolve something, it hands over with the reason and the conversation history, and the person picks up where the agent left off.
Workflow
A canvas for the work around the call
The workflow builder uses the same steps as the call flow. Pull a document, make a call, branch on the outcome, update the CRM, send a message, or schedule a follow-up. A step that needs a person - an approval - pauses the run until someone approves or rejects it. Nothing is committed by the workflow alone.
Pull and parse
Pull a document, extract the fields you asked for, and use them in the next step. A field that is not in the document comes back empty, not invented.
Make the call
Dial a contact that is already on a campaign, and carry the outcome forward.
Branch
Branch on what happened: resolved, booked, needs a person, or no answer.
Approval
A human approves or rejects before money, a booking or a message goes out.
Workflows live under /dashboard/workflows. A dry run lists the planned steps and sends nothing.
Call QA
Every call, scored from what was actually said
After a call, the agent scores itself from the recording and transcript: relevance to the caller's ask, whether it followed the playbook, whether its tool calls worked, and whether it invented anything. Scores that cannot be computed are shown as "no data", never invented.
Relevance and obedience
Did the agent answer what was asked, and did it follow the playbook and the guardrails? Scored only where the transcript gives clear evidence.
Tool-call success
Every tool the agent ran is logged with its result, so you can see whether bookings, lookups and transfers worked.
Hallucination flag
A call is flagged when the agent states a price, date or policy that is not in the knowledge base. The flag is only set on clear evidence.
Interruptions
Speaking over the customer is tracked where the call record allows it. Where the transcript has no timing data, the page says no data.
Live coach
During a live call, the human on the team sees suggested replies from the knowledge base, the next playbook step, and any guardrail note - so they can step in with context.
Improvements you approve
The system reads transcripts, finds what went wrong, and brings a fix to approve - a playbook change or a test. Nothing is auto-applied.
Guardrails
What the agent will and will not do
Grounded answers
The agent answers from your knowledge base only. If the document does not have the answer, it says so instead of inventing one.
Blocked topics
You set the topics the agent never takes on. Anything blocked is refused politely, and the refusal is logged.
PII handling
Card numbers, passwords and PINs are redacted, and the agent is told never to read them back. Identity is verified another way.
No invented prices or dates
The agent will not state a price, a date or a policy that is not in its knowledge base.
Transfer when unsure
When the agent is unsure, or the caller asks for something out of scope, it transfers to a person or takes a message - and logs the reason.
Approval gate
Any change to the prompt, flow, playbook or guardrails is staged and needs a human to approve it before it goes live.
Playbook
Greeting, resolve, wrap-up - every time
The playbook is the multi-step workflow the agent follows on every call: how it greets, what it does to resolve the call, which tools it may use, and how it wraps up with a summary. It is a structured set of steps, not a free-text prompt.
Greeting
The agent introduces itself, confirms who it is speaking to, and states what it can help with.
Resolve
The agent works the steps: check, look up, book, change, or transfer. Each step can call a tool.
Wrap-up
The agent summarises what was agreed, and the summary is pushed to your CRM or system of record.
Languages
English and Hindi today, Arabic where the voices are wired
Calls run in English and Hindi today. Arabic is wired where the voices exist: Microsoft Arabic voices and Chatterbox multilingual Arabic, including Gulf Arabic. We say Arabic is on the way elsewhere on the site only where the voice is actually connected - we do not claim a dialect or a language that is not wired.
Speech-to-text runs on Deepgram. Voices come from ElevenLabs, Cartesia, OpenAI and Microsoft. There is no voice cloning: the library is built-in voices only.
How it works
Live in four steps
1. Configure
Pick a template or write a prompt, set the purpose, and connect your carrier and model keys.
2. Choose voice
Pick a voice from the library and the language the agent should speak.
3. Build flow
Build the flow on the canvas, or generate it from a prompt, and check it in the playground.
4. Launch
Connect a phone number or the website widget, test a call, and go live. Monitor calls, QA, reliability and alerts after that.
Build at /dashboard/templates, test at /dashboard/settings/playground, deploy from /dashboard/telephony/numbers or /dashboard/widgets, monitor at /dashboard/calls/qa and /dashboard/settings/reliability.
Security
Governed on every turn
Tenant isolation
Every workspace is its own tenant. A call, a contact or a thread from another tenant never appears.
Sign-on and roles
Sign in with SSO, and give each team member a role: owner, admin, campaign manager or viewer. Permissions decide what each role can read, change or approve.
Encrypted credentials
Connector credentials - HubSpot, GoHighLevel, Salesforce, WhatsApp and the rest - are stored encrypted.
Webhook validation
Incoming WhatsApp, SMS and email webhooks are checked against their signature before anything is processed.
Audit log
Every agent change, approval, integration connection and post-call action is written to an audit log.
Recording and retention
Call recording and retention are configurable per workspace. You decide how long calls are kept.
We do not claim SOC 2, HIPAA, PCI or ISO here. If you need a specific assurance, ask us and we will tell you plainly what we have.
Deployment
Fits your environment
Deploy where your data, security and infrastructure requirements demand. Five of the six modes below run today; the sixth needs a control plane we have not built yet, and we say so.
Managed SaaS
Hosted. Fully managed, minimal operational overhead. From free to $299 a month.
Your cloud
Self-hosted on your own cloud or VPS. You run the same software we run, on infrastructure you control. $1,200 a year per domain.
Private cloud
Self-hosted, isolated infrastructure, complete control. Your server, your data, your models.
On-prem
Self-hosted on your own premises and server. The Docker Compose install runs the backend, the database and the local model services.
Air-gapped
Self-hosted with local models only: Ollama for the language model, faster-whisper for speech-to-text, and Kokoro for voices - all as local services in the install. Your licence is activated once by our team when you start.
Hybrid - not yet
Hosted for some workloads and self-hosted for others, under one workspace, is not built yet. It would need a shared control plane across both. Until then, each deployment is its own workspace.
Self-hosted is $1,200 a year per domain, 12-month term, licence required. If you do not renew, there is a 14-day grace period with a visible dashboard warning, then agents stop taking new calls. Calls already in progress are never cut off.
Integrations
Works with your stack
60 integrations in the catalogue. The agent can read or write your CRM, helpdesk, calendar and store during a call - look a ticket up, log a call, update a deal, book a meeting. Anything not listed connects through the generic API step, a webhook, or MCP.
CRM
HubSpot, GoHighLevel, Zoho, Salesforce, Pipedrive. Log calls, update contacts and deals.
Helpdesk
Zendesk, Freshdesk, Intercom, ServiceNow. Create and update tickets, and look a ticket status up.
Calendar and meetings
Google Calendar, Outlook Calendar, Calendly, Zoom. Check availability and book or move meetings.
Messaging
Slack, Microsoft Teams, WhatsApp, Telegram. Post call notes and alerts where your team works.
E-commerce and payments
Shopify, Stripe. Look an order up, and create a payment link so the caller can pay without reading card details to the agent.
Automation
Zapier, Make, n8n, generic webhooks. Connect anything with an API through the HTTP step.
Storage
Google Sheets, Microsoft Excel, Dropbox, Amazon S3, OneDrive. Export call logs and import contact lists.
AI and models
Bring your own keys: OpenAI, Anthropic, Google, Groq, Ollama. Your models, your bills.
Connect them yourself in the dashboard under /dashboard/integrations/connect, or on the integrations page. Every connection can be tested and disconnected.
Pricing
Flexible plans
designed to scale.
Choose a tier tailored to your outbound broadcasting and inbound calendar scheduling limits.
FAQ
Questions, answered plainly
Self-hosted or hosted - what is the difference?
Hosted runs on our infrastructure: free, $29, $99 or $299 a month, or $49 a month if you bring your own carrier and model keys. Self-hosted runs on your own server or VPS: $1,200 a year per domain, 12-month term, licence required. Self-hosted is not free.
Do the agents replace my team?
No. They handle the routine calls - booking, answering, qualifying - and hand the rest to a person, with the reason and the conversation history. After-hours they take a message instead of letting the call go unanswered.
What happens if a self-hosted licence expires?
There is a 14-day grace period with a visible dashboard warning, then agents stop taking new calls. Calls already in progress are never cut off.
Which languages do the agents speak?
English and Hindi today, and Arabic where the voices are wired (Microsoft Arabic voices and Chatterbox multilingual Arabic, including Gulf Arabic). We do not claim a language or dialect that is not wired.
Can the agent work on WhatsApp and email too?
Yes. It is the same agent - same prompt, tools and knowledge - on voice, WhatsApp, chat, web, SMS and email. WhatsApp uses the official Meta Cloud API, and outside the 24-hour window only an approved Meta template can be sent.
What does it cost?
Hosted: free, Starter $29, Growth $99, Scale $299 a month. Hosted BYOK: $49 a month, plus your carrier rates. Self-hosted: $1,200 a year per domain. There is no per-minute platform fee on BYOK.
Hear it handle your calls
Book a demo and we will walk you through a live agent on your own use case - or try a demo agent above and leave your number.