All pages

Behaviour

Tone, language, opening line, model and when to fetch a human.

Where to set it

Most of this is on the agent's Playground, in four cards: Language and model, Voice, What Agent says (the opening line and the instructions) and Reply length. Under them are the call settings: how the agent speaks, listens and ends a call, recording, memory, and call length (see Voice). Nothing changes until you press Save changes at the top right; after that it applies to the next message, including for real customers, so the Playground is not a sandbox. Test a call and a chat panel sit beside the cards (above them on a narrower screen), so you can try a change straight away.

Name

Rename an agent from the pencil beside its name at the top of its sidebar, or under Settings → Identity. Customers see the name in the chat.

Business description

The facts the agent falls back on when a knowledge search finds nothing — what you sell, to whom, and anything true of the business as a whole.

Keep it factual. This is grounding, not instructions: it is treated as information to work from, never as commands.

Instructions

Tone and subject matter. "Be warm but brief." "Never discuss competitors." "Always mention that delivery is free over ₹2,000."

Tune these in the Playground, against a live reply. Guessing at wording and shipping it is how you find out a week later that the agent has been ending every message with an emoji.

The box has no character limit — the counter under it shows how many characters you have used, and the text is sent to the model on every reply, so a longer prompt costs more on every turn. Through the API, instructions can be up to 20,000 characters. Conversation steps are limited separately (up to 12 steps, each with an 80-character title and a 600-character detail) and don't count toward the instructions.

Note

Your instructions are treated as guidance on tone and subject matter, not as a command channel. The platform's own safety rules always win — which is what stops a customer talking the agent out of them too.

Info

Your agent never pretends to be a person. Give it any name and persona you like, but if someone asks whether they are talking to an AI, a bot or a real person, it says it is an AI assistant — on every channel, in every language, whatever your instructions say. On a call, it also answers truthfully whether the call is being recorded.

Opening line

The first thing a visitor reads in the widget. Write it in the business's own voice; "How can I help you today?" is a wasted sentence.

Reply language

Choose it when you create the agent, or any time on the Playground. Either [Auto] Match the customer — the agent replies in whatever language the customer speaks in — or tick the languages it speaks: one, or up to four.

Through the API, send languages on Create Agent or Update Agent: a list of language names. [] is auto, one name is that language alone, and two to four are the languages the agent expects, the first being the one it falls back to, for example ["Telugu", "English"]. More than four, a repeat, a name the console does not offer, or languages together with language is refused with a 400. Get Agent returns languages too. Sending language (one language) still works and replaces any list.

In chat (website chat, WhatsApp, Telegram, the Playground and test chat), Match the customer reads the script of the customer's latest message. Write in Telugu, Hindi, Tamil, Kannada, Malayalam, Bengali, Gujarati, Punjabi or Odia script and the agent replies in that language, even if its opening line was in another. Once a conversation is under way, a few words in another script do not move it: the agent switches when the customer's first message is in that script, or when a later message is a real sentence in it. Some cases are left to the agent's own judgement: a message written in English letters (including Telugu or Hindi typed in English letters), a message too short to tell, and one that is mostly English. A customer can always ask for a different language and the agent follows.

Pick one language and the agent answers only in it. On a phone call that also tells the speech recogniser which language to listen for. Match the customer leaves the recogniser to work it out afresh for every sentence, and a short or mixed reply can be heard as the wrong language — a name from another language is enough. If your callers use one language, choose it.

Pick several — Hindi and English, say, or Tamil and English — and the agent answers in whichever of them the customer is speaking, and in the first one you ticked when they use any other. On a phone call the recogniser still works out the language for itself, but only among yours: a reply it hears as some other language is listened to again in the language the call is in. That second listen adds about half a second, and only to those replies. Up to four languages; the first one you tick is the one the agent falls back to.

On a phone call the agent is also told which of your languages each sentence the caller says is in, read from the script it was written in. A full sentence in another of your languages is answered in that language on the very next reply; a single word, a yes or a name keeps the language the call was in. When two of your languages share a script — Hindi and Marathi, say, or English and French — the script cannot tell them apart, so the agent is left to judge, as it is with Match the customer.

37 languages, including all 22 of the Eighth Schedule, searchable, each in its own script, so there is no "other languages" section to scroll past.

Speaking speed

In the Voice card, under the voice itself: how fast the agent talks on calls, on a slider from slower to faster. See Speaking speed.

Answer style

Presets rather than a raw temperature number. Precise for policy and pricing questions; warmer where the conversation is a sales one.

Model

Every new agent starts on Prana [Voice] — from scratch, from a pre-built agent, or through the API (prana-voice) — and answers chat and calls with it until you choose another. The choice is checked against your plan at request time, so downgrading a plan cannot leave a more expensive model quietly running.

If your workspace brings only its own voice key, calls run on Prana, Prana [Voice], GPT-OSS 120B or another low-cost model; an agent set to anything else answers its calls on GPT-OSS 120B. If it brings only its own language model key, your key answers on any model it offers, and the call is billed like any other call, by the voice. See Bring your own keys.

The model you choose never changes a call's per-minute rate: every call is billed at the premium rate, or the studio rate with a Studio voice. GET /api/v1/models gives the price for each model: callTier (always premium) and callPricePerMinuteMicro, the price of one call minute in millionths of currency (the currency of the wallet your calls are charged to). Both cover the model only, with a Premium voice: a Studio voice raises a call's price, and one call is billed at one rate, never a stacked surcharge. If your workspace brings its own keys, calls bill at the BYOK per-minute rate whatever the model, so these prices apply only to an agent you've set to run on our stack. The exception is a workspace that brings only its own language model: there the voice plan still applies, set by the voice alone, and the model you run never raises it.

Hover a model in the picker for a short card: who made it, what it suits, how it compares on intelligence and context, and how soon it starts speaking on a call in your region — the typical time (p50) and the time on slower turns (p95), measured on a real agent's calls. The number beside each model's name is that typical time.

Open models

Workspaces billed in rupees can also pick five open models, in chat and on calls:

ModelAPI idGood for
Gemma 4 26B A4Bgemma-4-26b-a4bThe quickest to answer; simple calls and FAQs
Gemma 4 31Bgemma-4-31bA little larger than Gemma 4 26B, a little slower
Nemotron 3.5 Super VL 120B A12Bnemotron-3.5-super-vl-120b-a12bThe largest here; long documents
DeepSeek V4.1 Flash 1Mdeepseek-v4.1-flash-1mQuick, with a 1M context for long documents
Qwen 3.8 27Bqwen-3.8-27bA mid-sized model

They answer text only for now: a photo a customer sends reaches them as a line saying a photo was sent. They're billed like any other model (GET /api/v1/models lists each one with its price), and they aren't offered to workspaces billed in dollars.

Models for workspaces billed in dollars

Workspaces billed in dollars can pick two more models, in chat and on calls:

ModelAPI idGood for
DeepSeek V4.1 Flashdeepseek-v4.1-flashQuick and steady on calls, with a long memory for documents
Nemotron 3.5 Lightningnemotron-3.5-lightningVery quick and light; short, simple calls

They aren't offered to workspaces billed in rupees, which have their own DeepSeek model above.

No longer offered: GLM-5.3 Flash (too slow to start speaking on calls) and GPT-4.1 Nano (retired by its maker). An agent that was set to either now answers on the default model; choose another in the picker.

Info

The agent does not go dark. If a model provider fails mid-conversation, the reply falls through to another model rather than to silence. A customer looking at a dead chat window is a worse outcome than a slightly slower answer.

Conversation budget

A ceiling on what one conversation can cost you in a day, under Settings → Conversation budget. When a single chat or WhatsApp conversation has cost that much in the last 24 hours, the agent keeps answering on a lighter model until the day is over. It never stops replying. Leave it blank for no budget; it is in your balance's currency.

It is for the visitor who will not stop typing, or a loop between two bots. A regular customer's thread is counted a day at a time, so they are never moved to the lighter model for good.

Handing over to a person

One switch, Actions → Hand over to a person, with two triggers behind it — and Escalation covers what happens when one fires, including when a custom action fails:

  • When the material does not cover the question. The honest escalation.
  • When the customer asks for a person, or is clearly upset.

Turning it off removes the capability entirely rather than asking the model to refrain — the difference matters, because a model that has a tool will eventually use it.

Tip

A business with nobody watching the inbox should turn escalation off. An escalation nobody answers is worse than a straight "I don't know", because the customer waits.

Writing your own lead and hand-over rules

Every agent is told, in plain sentences, when to ask for contact details, when to hand a conversation to a person and what to say when it cannot. If you build on the API and write those rules yourself, in your own instructions or scripts, set ownRules: true on Create Agent or Update Agent and we leave ours out.

What goes: the sentence about asking for a name, email or number (and the one that says the business does not collect them), the hand-over rule (and the one that says nobody can be handed to and that nobody should promise a follow-up), and the "cannot arrange a call back" sentence. What stays, whatever you send: the boundary (only this business), the AI disclosure, the reply language, the date, your own instructions, and how a call is put through to a person.

It is off by default, so no agent's behaviour changes until you set it. The tool that hands a conversation over still describes when it applies, so an agent keeps handing over when someone is on duty. captureLeads and the escalation switches keep deciding which tools the agent has; ownRules only decides which of our sentences it reads.