Five panels and a test call
No-code voice agents, written rather than drawn
There is no flowchart to draw and nothing to deploy. You describe the job in sentences, upload the documents it should answer from, tick the things it may do, and call it yourself from the browser before a single customer reaches it.
Everything you fill in, in order
Panel one
Say who the agent is
The behaviour of the agent is a piece of writing, not a diagram. You describe the job in ordinary sentences — who it is answering for, what it may promise, when it should stop and fetch a person — and that is the whole configuration.
- Write the greeting the way you would say it
- Set the tone, and what it must never offer
- Say what happens outside opening hours
Panel two
Give it your answers to read from
Paste text, point it at your website, or upload PDFs — fee lists, timings, policies, directions. The agent searches that material during the call and answers from it, instead of improvising.
- Upload a document or paste a page of text
- Add the website addresses worth reading
- Re-upload whenever the price list changes
Panel three
Pick the voice and the languages
Speech recognition, the language model and the voice are each chosen from a list on the agent’s pages. Recognition covers 23 languages with automatic detection, so a caller can start in Hindi and finish in English; the voice you reply in is picked per agent, with Indian-language voices in 11 languages.
- Choose a voice and hear it
- Tune the speech controls that voice exposes
- Run a second agent in a second language if you need one
Panel four
Switch on what it may do
Tools are turned on by name from a catalogue: check an order, find a free calendar slot, book it, update a contact, send a WhatsApp message, call one of your own web endpoints. Nothing here is code — connecting the account behind a tool is a settings screen.
- Tick the tools the job needs
- Connect the calendar, store or inbox behind them
- Leave the rest off
Panel five
Call it yourself, then give it a number
Talk to the agent from your browser before any customer does, read the transcript of your own call, and fix the wording you did not like. Only then do you point a phone number at it — one you rent in the console after a KYC check, or one you already own through your own carrier.
- Place a test call from the browser
- Read your own transcript and edit the prompt
- Forward your line, or import your existing number
Wire one up here, then ring it
The same pieces as the five panels above, as switches. Tick what the job needs and read the agent it makes back.
The switches
Each one is a real panel in the console.
Your agent so far
It answers the phone and answers for the business the way you described, replies from your uploaded documents and speaks in the voice you chose.
Panels covered
- Panel one — Say who the agent iscovered
- Panel two — Give it your answers to read fromcovered
- Panel three — Pick the voice and the languagescovered
- Panel four — Switch on what it may donot yet
- Panel five — Call it yourself, then give it a numbernot yet
Simulated example — the real console has more switches, but these are all you need to start.
Then call it yourself
Simulated example
Agent: Good afternoon — thanks for calling. How can I help?
You: Do you have anything free this Friday?
Agent: Friday has eleven thirty or four fifteen. Which would you like?
You: Four fifteen.
Agent: Booked for Friday at 4:15 — and I have sent the address to this number on WhatsApp.
Your own test call leaves a transcript like this to read back — fix the wording you did not like, then call it again.
Why there is no flowchart to draw
Older phone bots were drawn. Every sentence a caller might say had to exist as a branch before they said it, which is why a stray word ended in “I did not understand that”. A written instruction does not have that shape.
- How is behaviour defined?
- A drawn phone treeBoxes and arrows. Every path a caller might take has to have been drawn in advance.
- A written instructionSentences. You describe the job; the model handles the wording the caller actually used.
- A caller says something unplanned
- A drawn phone treeFalls off the diagram — a default branch, or "I did not understand that".
- A written instructionAnswered from your documents, or escalated with a written brief if it is genuinely out of scope.
- Changing your mind on a Friday
- A drawn phone treeRe-draw the branch, re-record the audio, republish the tree.
- A written instructionEdit the paragraph, place a test call, keep the old version in history in case you preferred it.
- Knowing whether the change helped
- A drawn phone treeButton-press counts.
- A written instructionReplay a suite of scripted callers against the change, or split live traffic between two versions and compare.
It is Friday afternoon and the greeting has to changeSimulated example
- Edit the paragraph
- Place a test call
- Keep the old version in history in case you preferred it
If you do still want a keypad somewhere — an accessibility option, a legal disclosure branch — a call flow can hold one and route it to an agent. That path is set out on replacing an IVR.
Nothing here is a dead end
The same agent has an API
Nothing is rebuilt when a developer finally joins. The agent you configured in the console is reachable programmatically, so code is added where it earns its place rather than at the start.
Your own endpoints become tools
A REST endpoint you already run can be attached as a tool with a description of what it does. An MCP server can be attached the same way.
Several agents behind one number
Group a billing agent, a support agent and a booking agent into a squad behind a single entry point, with a policy that stops a caller bouncing between them.
The developer platform · Documentation · What connects without code
Before you open the builder
What are no-code voice agents?
They are AI phone agents you build by filling in a form rather than by programming. You write what the agent should do in plain sentences, upload the documents it should answer from, tick the tools it may use, pick a voice, and point a phone number at it. There is no flowchart to draw and no deployment to run.
Do I genuinely need no developer?
For a standard deployment, no. What you do need is the content the agent answers from, and access to your phone forwarding settings — or a number rented from the console after a KYC check. A developer becomes useful only when you want to attach an internal system that has no ready-made connector.
Is a no-code agent a weaker agent?
No. No-code describes how you configure it, not what runs the call. The same voice pipeline handles a console-built agent and an API-built one: streaming speech recognition, the same model choices, the same interruption handling, the same tools.
Can I hear it before customers do?
Yes, and you should. Place a call to your own agent from the browser, then read your transcript back and edit whatever sounded wrong. Nothing is live until you forward a number to it.
What if I change the prompt and make it worse?
Prompt versions are kept, so you can compare and go back. You can also replay a suite of scripted callers against a change before it ships, and split live traffic between two versions to see which one does better on a measure you choose.
What happens when the agent cannot answer?
It tells the caller someone will ring them back and files a handoff with a short written brief for whoever picks it up. Your team returns the call from the handoff inbox. It does not connect a live caller straight through to a person, so plan around callbacks.
Can I start without paying?
Yes. A new account opens with free credits, which is enough to build an agent and call it yourself from the browser several times over. You pay from prepaid credits for the minutes the agent spends on real calls, and credit packs start at ₹99.
Open the console and write the first sentence
Where these agents end up: answering the calls you receive, working a list you upload, covering a single line after closing and what the minutes cost.