AI Agents

Open the AI Agents page in Callin.io The AI Agents page is where you create and manage the agents that handle your phone calls and website conversations. From here you can create a new agent, find an existing one, test it, and open its editor. Each agent’s editor is organized in nine tabs at the top: Agent, Workflow, Knowledge Base, Conversation, Analytics, Tools, Call Action, Website widget, and Advanced. This page explains how to create an agent and what you can do in each tab. Everything starts on the AI Agents page, where each row is one of your agents:

​

The agents list

Every agent in your account appears as a row with its name, the phone number assigned to it (or No phone number assigned), its call direction, its language, and the date it was created.

• Search agents… — find an agent by name.
• Direction — show only Inbound, Outbound, or Both agents.
• Sort — order the list by Newest first, Oldest first, Name A–Z, Name Z–A, or Custom order.

On the right of each row you have three quick actions:

• Test Call — place a test call with the agent.
• Test AI agent — talk to the agent from your browser.
• More options (⋮) — Copy ID, Clone, Edit, Assign Number, View events calendar, and Delete.

​

Create a new agent

Click Create Agent at the top right. Callin.io asks how you want to start. Start from Scratch builds an agent with a basic prompt and greeting that you then shape in the editor. Choose it when you already know what the agent should do. You only need to enter the Agent Name (up to 50 characters) and choose the Agent stack. Browse our Templates starts from a ready-made agent designed for a specific business and use case. It’s the fastest way to get a working first version: Predefined Templates (Beta) uses templates created by your own organization, so every new agent starts with your company’s standard setup. Each card shows the template’s name, a short description, and how many fields it pre-fills (for example Agent Name, Prompt, Greeting Message, Direction). Select one, review the Agent Name, choose the Agent stack, and click Create Agent. If your organization hasn’t created any templates yet, this list is empty.

​

Choose the agent stack

The Agent stack decides which voice and AI technology powers the agent:

• Premium Voice — natural Callin.io voices, ready to deploy in minutes. This is the default and the right choice for most agents.
• Custom Stack (Alpha) — bring your own LLM, TTS, and STT for full control over each part of the voice pipeline.
• Ultra-low latency — near-instant responses for smooth, live conversations. These agents have their own editor tabs and are the only ones that can join a squad. See Ultra-low latency agents.

Click Create Agent to finish. The agent opens directly in its editor, on the Agent tab.

​

Saving your changes

When you change anything in the editor, a bar appears at the bottom with the message You have unsaved changes:

• Save — stores all your changes.
• Clear — discards every unsaved change and restores the last saved version.

While there are unsaved changes, Test Call is disabled. If something needs fixing before saving (for example, an incomplete workflow), the bar tells you what’s missing or shows a View Errors button.

​

Testing your agent

Use the buttons at the top right of the editor to test without leaving the page:

• Test Call — places a real call with the agent.
• Test AI agent — lets you talk to the agent from your browser.
• Console — opens the developer console.

​

Section 1: Agent

The Agent tab is where you define the core identity and behavior of your assistant: what it says, which voice and language it uses, which AI model powers it, and whether it answers incoming calls, makes outgoing ones, or both.

​

Conversation Prompt

The Conversation Prompt is the main set of instructions your agent follows during a call: the goals it must achieve and the guardrails it must respect to stay on-brand and compliant. Think of it as the agent’s playbook. A new agent starts with a template organized in three sections you should adapt:

• Personality — who the agent is and the role it plays.
• Environment — the context it operates in (support call, appointment booking, reception, and so on).
• Tone — how it should sound (calm, friendly, professional, brief).

After those three, add operational rules that guide the agent’s decisions during the call. Some that work well for most setups:

• Ask only one clarifying question at a time.
• Always move the user toward the next useful step.
• Use the knowledge base for reference answers.
• Keep responses short unless the user asks for detail.
• Don’t invent answers. If the information is unclear, say so.

​

Add Variables and Dynamic Variables

The Add Variables button below the prompt editor (and below the greeting) inserts placeholders that are replaced with real values during the call. The menu has two groups:

• System Variables — values Callin.io fills in automatically: system__agent_id, system__called_number, system__time_utc, system__time, system__timezone, and system__call_id. Ultra-low latency agents only have system__time and system__timezone.
• Custom Variables — your own variables, such as agent_name. Click New Variable to create one.

Variables appear as chips below the editor and in the Dynamic Variables panel at the bottom of the tab, where you set their value. For example, {{agent_name}} is set to the agent’s name by default. This is the right approach when you want to reuse the same agent for different brands, locations, or services without rewriting the prompt every time.

​

Time zone

Next to Add Variables there’s a time zone selector (for example, Rome (GMT+1)), with a search box to find any city. This is the time reference the agent uses when it talks about today, tomorrow, this afternoon, or any specific time during the call. Always set it to the real time zone of the business the agent represents. A wrong time zone leads to wrong appointment times.

​

Default personality

The gear icon next to the time zone opens the Default personality toggle. It’s off by default. When you turn it on, Callin.io adds a set of general guidelines to the prompt, for example: don’t repeat who you are unless asked, give helpful and informative answers, ask clarifying questions when needed, and be polite, professional, clear, and concise. Turn it on if your prompt is short and you want these good practices added for you. If your prompt already covers them, leave it off.

​

Greeting Message

The Greeting Message is the very first thing the agent says when the call starts. A good greeting tells the caller who they’re talking to and what kind of help is available, in one short sentence. You can use variables here too, for example Hi, this is {{agent_name}}—thanks for calling, what can I do for you today? A few examples:

• Support: “Hi, I’m the support assistant. Tell me what’s going on and I’ll help you step by step.”
• Appointments: “Hi, I’m the booking assistant. I can help you schedule, change, or cancel an appointment.”
• Reception: “Hi, thanks for calling. I’m the virtual assistant and I’ll help you with your request.”

​

Allow Interruptions

Allow Interruptions, next to the greeting, lets callers interrupt the agent while it’s delivering the first message. It’s on by default, which feels natural for most support, reception, and customer service flows.

​

Voice

Voice controls how your agent sounds. The voice doesn’t change what the agent says, but it heavily affects how callers feel during the call — clarity, trust, and overall impression. Click the selected voice (marked Primary) to open the voice panel:

• Voice — click it to open Select Voice, which has three tabs:
  • Explore — the catalog of available voices, with a search box and filters by language and gender.
  • Cloned — voices you’ve cloned in your account.
  • Add — import a voice from ElevenLabs by pasting its VoiceLab share link and giving it a name.
• Expressive Mode — on by default. Lets the agent express speech with audio tags, like an emotion, a speed, or an accent.
• Suggested audio tags — up to 20 tags the voice can use for emotional nuance, such as Excited, Concerned, Patient, Whispering, US accent, or Laughing. Click a tag to add it and describe when the agent should use it, or click Add to create your own. The tags are suggestions: the agent can also use tags not listed.

​

Add additional voice

Click Add additional voice to attach secondary voices, useful when one agent needs to handle several languages or different customer segments. For each one you set:

• Voice — the voice to use.
• Label — a unique identifier the model uses to refer to this voice.
• Language — the language for this voice (by default, Same as agent).
• Description — when the agent should use this voice, for example “For any Spanish words” or “Whenever Joe is speaking.”

Click Add voice to save it.

​

Common voice settings

The gear icon next to Voice opens Common voice settings, two audio controls shared by all the agent’s voices:

• TTS output format — the format of the generated audio: PCM 8000 Hz, PCM 16000 Hz (Recommended), PCM 22050 Hz, PCM 24000 Hz, PCM 44100 Hz, PCM 48000 Hz, or ulaw 8000 Hz (Telephony). PCM 16000 Hz is a good balance between audio quality and processing speed.
• Optimize streaming latency — a scale from 0 to 4 (default 3). Higher values make the agent start speaking faster, at a slight cost in audio quality.

​

Language

Language sets the agent’s default language (marked Default). Click it to choose another one from the searchable list. Click Add Additional Languages to pick the other languages the agent can switch to during the conversation. The panel has a search box and a Select all languages option. Only enable the languages you actually need. Adding extra languages “just in case” tends to make behavior less predictable. To switch languages automatically, the Language Switching system tool must also be on (see Section 6: Tools).

​

AI Intelligence (LLM)

AI Intelligence (LLM) is the language model that powers the agent. The LLM (Large Language Model) is what understands the caller, follows the rules in your prompt, uses the knowledge base, and generates each response. Click the selected model to open the LLM panel, and click it again to see the list. Each model shows a short description and its approximate latency, like ~179ms. The models available include GPT-OSS 120B, GLM-4.5 Air, GPT-4.1 Nano, GPT-5 Nano, Gemini 2.5 Flash Lite, and Gemini 2.5 Flash (the default). The model directly affects three things:

• How well the agent understands what the caller means.
• How fast it answers.
• How consistent it stays across long conversations.

For appointment-focused agents and scheduling flows, Gemini 2.5 is a solid practical choice.

​

Thinking Budget and Temperature

The LLM panel also includes two settings:

• Thinking Budget — controls how many internal reasoning tokens the model can use before responding. Off by default. More reasoning can improve complex answers but adds latency, so test it on real calls.
• Temperature — controls how creative or deterministic the responses are (default 0.5). Lower values make the agent more focused and predictable; higher values make it more creative and less predictable.

For support, appointments, and operational agents, keep temperature low. Save higher temperatures for creative or exploratory use cases.

​

Call Direction

Call Direction decides how the agent participates in the calling flow:

• Outbound — the agent makes outgoing calls (campaigns, reminders, confirmations, follow-ups). This is the default for a new agent.
• Inbound — the agent answers incoming calls (support, reception, customer service).
• Both — the same agent works for incoming and outgoing calls.

​

Section 2: Workflow

The Workflow tab is where you decide how the conversation moves. While the Agent tab defines who the assistant is, Workflow defines the path the call follows: branches based on what the caller wants, transfers, tool calls, and clean endings. You build it on a visual canvas by placing nodes and connecting them. Use the buttons at the top left to zoom in, zoom out, and fit the whole flow on screen, and the buttons at the top right to go full screen or delete the workflow.

​

What you can build here

From a single starting point, you can:

• Create multiple routes inside the same conversation.
• Send the call to a Subagent for a specialized stage.
• Transfer the call to another agent.
• Transfer the call to a phone number.
• Run a tool that already exists.
• End the conversation with an End node.
• Control which route applies using LLM conditions.

​

Adding nodes

To add a node, drag a connection from an existing node (for example, from the bottom of Start) and drop it on an empty part of the canvas. You can also click the + button below a selected node. A menu opens with the node types: Subagent, Agent transfer, Phone number transfer, Tool, and End. The new node appears connected to the previous one, and its settings open in the panel on the right.

​

Start

Start is the entry point. Every conversation begins here. The node doesn’t speak or act — it just launches the route and sends the call to the next node. The side panel for Start has two tabs:

• Overview — explains its role.
• Edges — lists every outgoing route from Start, with its label and destination.

​

Transitions and LLM conditions

A transition is the line that connects one node to the next. Click its label (at first, Configure condition) to open its settings. The Forward tab defines the transition:

• Transition type — LLM Condition (the model decides when to take this route), None, or Expression (not yet editable in this panel).
• Label — the name shown on the canvas.
• LLM condition — a short text that tells the model when this route should apply.

Examples of conditions:

• “Use this route when the user needs technical help with the platform.”
• “Use this route when the user asks to speak with a person.”
• “Use this route when the issue is fully resolved and the conversation should end.”

​

Complete your workflow setup

The Complete your workflow setup panel lists exactly what’s missing before the workflow can be saved. Each item shows the node or edge to fix and what it needs. Until the list is empty, Save stays disabled. Common items it flags:

• An Agent transfer node without a target agent.
• A Phone number transfer node without a destination type.
• A Tool node without a tool selected.
• A transition with no LLM condition text.

​

Subagent

A Subagent is a specialized stage inside the same workflow. Use it when one part of the conversation needs its own goal, voice, model, or behavior, different from the main agent. Give it a name at the top of the panel, then configure its tabs:

• General — Conversation Goal (instructions that extend the main prompt for this stage), Override prompt (to replace the main prompt instead of extending it), Voice, LLM, Eagerness, and Spelling patience.
• Knowledge Base — Inherit knowledge base is on by default. Turn it off to use a different one.
• Tools — Inherit custom tools is on by default. Use Add Tool to attach specific tools.
• Edges — the routes that leave this node.

Tools themselves are not created here. They’re created in the global Tools section and only selected or inherited from inside the workflow.

​

Tool node

The Tool node (shown as Tool dispatch in the side panel) sends the conversation to a stage where the agent uses an existing tool. Use Add Tool in its Tools tab to select at least one. This node doesn’t create new tools — it just points the workflow to tools that already exist.

​

Agent transfer

Agent transfer hands the conversation over to another agent in your account. Use it when the call should move to an assistant specialized in another task, language, or workflow. You can configure:

• Agent — the target agent.
• Delay (ms) — how long to wait before the transfer, in milliseconds.
• Transfer message — an optional message the agent says during the transfer.
• Enable first message — whether the target agent plays its greeting after the transfer.

Examples of transfer messages:

• “I’m going to transfer you to the right agent for this request.”
• “I’ll connect you with the assistant specialized in this type of question.”

​

Phone number transfer

Phone number transfer sends the call out to a phone number. Use it when the conversation has to leave the agent and continue on a specific phone line. The Destination type can be:

• Phone number — a fixed number, ideal for reception, a sales line, or a dedicated support line.
• Phone (dynamic variable) — the destination comes from a dynamic variable decided at runtime.

​

End

The End node closes the conversation: the call ends when this node is reached. Use it when:

• The issue is fully resolved.
• A branch should close cleanly.
• A stage is complete and shouldn’t continue.

​

Duplicate and delete

When you select a node, two icons appear next to it to duplicate or delete it. Duplicate is handy for building similar branches faster. Before deleting, check the transitions connected to that node so you don’t break a live route.

​

Global settings

When no node is selected, the right panel shows Global settings. Prevent infinite loops stops the workflow from continuously transitioning in a loop when all conditions are true. To disable a workflow without deleting it, disconnect the Start node. Without a connected start, the flow has no entry point and won’t run.

​

Section 3: Knowledge Base

The Knowledge Base tab gives your agent extra background information to answer questions. It’s best for large or detailed content — like manuals, FAQs, or policies — that doesn’t fit in the main prompt.

​

What this tab is for

This tab connects documents from your global Knowledge Base to a specific agent. It does not create content — that’s done in the Knowledge Base manager. Quick access:

• Open the Knowledge Base manager

A useful way to think about it:

• The prompt defines how the agent should behave.
• The knowledge base defines the reference content the agent can pull from when answering.

​

When to use a knowledge base

Use it when the content is:

• Long (manuals, policies, internal documentation).
• Frequently updated.
• Already published as articles or pages.
• Better kept separate from the prompt for easier maintenance.

Typical sources: FAQs, product guides, support articles, policy documents, internal references.

​

Supported content sources

The Knowledge Base manager organizes content in two types of sources:

• Files — uploaded documents.
• Web — published pages or online documentation.

​

What you see when a base is linked

The tab lists every document the agent can use. If there are none yet, it shows No documents added.

​

Main actions

Click Add Document to open the document picker:

• Use the search box to find a document from your Knowledge Base and select it to make it available to this agent.
• If the document doesn’t exist yet, click Add Document at the bottom of the picker. It takes you to the Knowledge Base manager, where you can create it.

​

Select Knowledge Base panel

In the Knowledge Base manager you can switch between Files and Web, search your content, and use Create Knowledge base to organize documents into groups. Once a document is there, it appears in the agent’s document picker.

​

Recommended flow

The cleanest way to work is to keep creation and selection separate:

​

Section 4: Conversation

The Conversation tab lists every call handled by this agent, incoming and outgoing. It’s where you review what happened on each call. It has two sub-tabs: Conversations (phone and web calls) and Google Meet Sessions.

​

Search and filters

• Search transcripts… — find calls by what was said during the conversation.
• Filters — narrow the list by Phone Number, Call ID, Squad, Status, Direction, Appointment, Call Transfer, Voicemail, Date Range, Call Duration, Origin, Campaign ID, Sentiment, or Ticket ID.

Until the agent makes or receives its first call, the tab shows No incoming or outgoing calls yet.

​

Google Meet Sessions

This sub-tab lists the Google Meet meetings the agent has joined (with the Join Meet button at the top of the editor). Search by the meeting URL or join code, and filter the list by Status. If the agent hasn’t joined any meeting, you’ll see No Google Meet sessions found.

​

Section 5: Analytics

The Analytics tab shows how this agent is performing, compared with the previous period of the same length.

​

Time range and grouping

• Time range — Last 24 hours, Last 7 days (default), Last 30 days, Last 90 days, or Custom range.
• Group by — Hour, 4 hours, Day, Week, or Month.

If there are no calls in the selected period, the tab suggests a wider time range or a test call, and offers a View conversations button that takes you to the Conversation tab.

​

Section 6: Tools

The Tools tab decides what your agent can actually do during a call beyond having a conversation. While the prompt defines how it talks and the workflow defines how the call moves, tools let the agent take action at the right moment. The tab has two sub-tabs: Tools, for your custom tools and the built-in System Tools, and MCP, for MCP servers. Quick access:

• Open the global Tools section

​

Custom tools

Custom tools are tools your team builds for specific use cases. They’re not created inside the agent — first you build them in the global Tools section, and then you come back here to attach them. Typical things a custom tool can do:

• Query an external system.
• Run an automation.
• Send or receive data through a webhook (a URL your tool calls when something happens).
• Trigger business-specific logic.
• Connect the agent to another service.

The Add Tool button opens a picker where you can search your tools and select one. If your account has no tools yet, the picker offers Add Webhook Tool to create one.

​

MCP servers

The MCP sub-tab lets you connect MCP (Model Context Protocol) servers to give your agent additional tools and integrations without building them one by one. Click Add MCP to search the MCP servers available in your account and connect one to the agent. If there are none yet, click Manage MCPs to set them up.

​

System Tools

System Tools are built-in tools that handle common call situations without you having to build anything. Each one can be turned on or off. A new agent starts with three of them on — Smart Hang-up, Language Switching, and Voicemail Intelligence — and the counter at the top of the list shows how many are active. Active tools with extra settings show a gear icon. Every tool’s settings panel includes:

• Description (optional) — explains to the model how and when to use the tool. Leave it blank to use the default optimized prompt.
• Disable interruption — prevents the caller from interrupting while the tool is running.

​

Smart Hang-up

Lets the agent politely end the call once the conversation goal is met or the caller says goodbye. Its description already includes examples of explicit endings (“bye”, “that’s all”, “no thanks”) and implicit ones (minimal answers once the caller’s needs are met, “I need to go”). This avoids keeping the line open longer than needed.

​

Language Switching

Lets the agent recognize and switch to the caller’s preferred language during the call. It only works with the languages you added under Language in the Agent tab.

​

Intelligent Listening

Keeps the agent quiet when the caller says things like “give me a second” or “hold on”, preventing awkward interruptions while they check something.

​

Transfer to agent

Hands the call off to a more specialized AI agent when the caller’s request falls outside the current agent’s expertise — for example, from a generic assistant to a sales agent. When you turn it on, its panel includes Transfer Rules. Click Add Rule to define each one:

• Agent — the agent to transfer to.
• Condition — when this transfer should happen.
• Delay before transfer (milliseconds) — how long to wait before transferring.
• Transfer message — the exact phrase spoken before transferring.
• Enable First Message — whether the target agent plays its greeting after the transfer.

​

Transfer to number

Connects the caller with a live team member when a human touch is needed or the caller asks for it — reception, a switchboard, sales, or a specific office.

​

Human Transfer Rules

The Transfer to number panel includes Human Transfer Rules, where you define the conditions for transferring to human operators. Click Add Rule for each one:

• Destination Type — a fixed Phone Number or a number provided through a dynamic variable.
• Phone Number — the number to transfer to.
• Condition — when this rule should apply.

Use Remove Rule to delete one. Use a fixed number when the destination is always the same, and a dynamic variable when it depends on the conversation context.

​

Voicemail Intelligence

Recognizes when a call reaches an automated inbox and acts accordingly. This is critical for outbound agents, where calls often end up at voicemail. In its settings, Voicemail Message (optional) decides what happens:

• Leave it blank to end the call immediately when voicemail is detected.
• Write a message (for example, “Hello, this is an automated call. Please call us back at your convenience. Thank you.”) to play it before ending the call.

The gear icon only appears while Voicemail Intelligence is on. Click it to open Edit Voicemail Detection Tool, where you’ll find Voicemail Message.

​

Access Property Data

Lets the agent use configured HTTPS XML feed URLs with property listings during the conversation. It’s designed for real estate agents that need to answer questions about available properties. It’s off by default.

​

Description and Disable interruption

As mentioned above, every system tool’s panel includes Description and Disable interruption. Write a description only when you need the tool to behave differently from the default — for example, to tell Smart Hang-up not to end the call until a booking is confirmed.

​

Choosing the right tools

Don’t enable everything. Pick what fits your use case:

• An inbound support agent often uses Smart Hang-up, Language Switching, and Intelligent Listening.
• An agent that escalates to humans uses Transfer to number with clear transfer rules.
• An outbound agent often uses Voicemail Intelligence, Smart Hang-up, and Language Switching.

​

Section 7: Call Action

The Call Action tab is where the agent stops being just a voice and starts producing useful outcomes: bookings, follow-up emails, tickets, structured data, and synced records — during and after the call. Quick access:

• Set up Google Calendar
• Set up Cal.com

​

Appointment Scheduling

Lets the agent book meetings directly into your calendar during the call. Especially useful for sales, onboarding, and any flow where booking on the spot saves you from chasing the customer afterwards. When you turn it on, the Scheduling Configuration panel asks for a Scheduling Provider:

• Google Calendar — best when you want to book directly into a connected Google calendar.
• Cal.com — best when your booking operation already runs on Cal.com.

Then select the calendar and the booking template the agent should use. You can create several booking templates per calendar and assign templates from different calendars to different agents.

​

Post-Call Summary Email

Automatically sends the call details to the email addresses you choose as soon as the conversation ends, so owners or teams stay in the loop without manual follow-up. When you turn it on, the Post Call Email Configuration panel opens. Choose a template type:

• Default template — fill in Recipients (type an address and press Enter or click Add), Subject, and Message.
• Custom template — write the whole email in HTML in Template body. Use Add Variables to insert call data such as userName, phoneNumber, direction, agent_name, duration, extractedDataHtml, keyMomentsHtml, transcriptHtml, and rawDataHtml. Format HTML tidies the code and Preview Email shows the result.

You can also set:

• Email Language — the language of the email.
• Include Call Data — adds the call data to the email.
• Include Extracted Fields — adds the data captured with Information Capture (see below).

Once configured, a gear icon next to the toggle reopens the panel.

​

Smart Opt-Out

Gives callers a polite way to decline future contact. Important for outbound flows: it keeps your lead lists clean. On its own, it doesn’t make your campaigns compliant with the rules of your country.

​

Issue and Complaint Tracking

Detects issues or complaints during the call and automatically creates a ticket, so your team can review, follow up, or escalate it from the Tickets section. Turn it on when you want each conversation to be able to create a case worth handling afterwards.

​

Data Sync (Post-Call Webhook)

Sends the call details and captured data to another system the moment the call ends. The right choice when you work with a CRM, an automation platform, or any internal tool that needs the data without manual copy-paste. To set it up, enter the webhook URL that should receive the data, and confirm in advance which system will receive it and how it will process it.

​

Summary Language

Sets the language of the conversation summaries. The default, Auto (infer from conversation), writes each summary in the language the call was held in. Pick a fixed language from the searchable list if your team always needs summaries in the same one.

​

Information Capture (Data Extraction)

Turns the conversation into structured data — names, dates, preferences, cities, products of interest, contact reasons, and so on. Click Add Entity to define each piece of data to capture. The Configure Data Extraction panel asks for:

• Data Type — String, Boolean, Number, or Integer.
• Identifier — the internal name used to store it.
• Description — exactly what the system should extract. It’s passed to the model, so describe in detail how to find it in the transcript.
• Enum Values (optional) — a closed list of values the model can choose from. If you don’t add any, the model can use any value.

The counter above the button shows how many entities you’ve defined.

​

Capture and Store Contact Data

When this is on, the extracted information gets saved into the contact’s profile. Useful when you want to reuse this data on future calls or keep richer records of each contact.

​

Call Retry

Persistence & Connection (Call Retry) sets up automatic retries when a call reaches voicemail. Mostly used in outbound agents. You can configure:

• Retry Attempts (Retry Count) — how many times the system should try again before stopping (default 0, no retries).
• Wait Time Between Tries (Retry Delay) — how many minutes to wait between attempts (default 30).

​

Section 8: Website widget

The Website widget tab lets you publish your agent on a web page. Visitors can talk to it by voice, by chat, or both, directly from your site, landing page, support page, or client portal. This tab combines two things: the technical embed snippet, and the visual and textual customization of the experience.

​

Embed snippet

The Embed snippet is the code fragment you copy into your site’s HTML to make the widget appear. It already includes your agent’s ID. The Copy button copies it so you can paste it directly. Where to paste the snippet depends on the widget mode:

• Inline embed — paste it exactly where you want the widget to render on the page.
• Floating launcher — paste it near the end of the page body so it sits as a floating entry point.

​

Layout & behavior

This block decides how the widget gets embedded and how visitors interact with it.

​

Widget mode

How the widget appears on your site:

• Inline embed — renders the full widget directly inside the page, where you placed the snippet.
• Floating launcher — shows a small orb that opens the widget in a sheet when clicked.

​

Embed widget theme

Controls the overall look:

• Match app theme — follows whatever theme your site uses.
• Light — light mode.
• Dark — dark mode.

Pick the option that fits your site’s design. Readability and brand consistency come first.

​

Conversation mode

How visitors can interact:

• Voice only — voice interaction only.
• Chat only — text only.
• Voice + Chat — both.

​

Appearance

Controls the visual style of the dial and the launcher.

​

Widget variant and dial size

• Widget variant — base visual size of the widget: Tiny, Compact, or Full.
• Dial size (px) — the minimum and maximum diameter of the dial, in pixels.

​

Realtime transcript and mute button

• Realtime transcript — shows a live text transcript while the conversation is happening.
• Show mute button — lets visitors mute and unmute their microphone during the call.

​

Orb colors

• Orb primary color and Orb secondary color — the main colors of the widget.

Use brand colors but keep enough contrast for readability.

​

Dial labels

Texts shown in the center of the dial depending on the conversation state:

• Center (idle)
• Center (connecting)
• Center (listening)
• Center (interrupt)

​

Text labels

Override some important interface texts. Any you leave empty use the default:

• Main label
• Start call
• Start chat
• End call
• Chatting status
• Placeholder (new conversation)

​

Chat UI labels

Texts specific to the chat experience:

• Voice tab label
• Chat tab label
• Chat title
• Chat subtitle
• Chat status (idle), Chat status (connecting), and Chat status (live)
• Chat start label
• Chat end label
• Chat empty state
• Input placeholder (idle) and Input placeholder (active)
• Send label

​

Adapting the widget to your use case

A few combinations that work well:

• Commercial landing page — Inline embed, light or matched theme, Voice + Chat, Full variant, contact-oriented copy.
• Website support — Floating launcher, Chat only or Voice + Chat, Compact variant, help-oriented text.
• Voice-first experience — Inline embed, Voice only, larger dial, visible mute button, short and direct labels.

​

Section 9: Advanced

The Advanced tab is where you fine-tune the agent: how it captures audio, how it handles silence and waiting times, how it manages call data, and how it performs in outbound telephony scenarios. Most agents work fine without touching anything here. Open this tab when you want more precision, or when something specific isn’t quite working. The tab is split into seven blocks: Memory, Voice & Understanding, Conversation Flow, Background sound, Soft timeout, Privacy & Data, and Telephony Extras.

​

Memory

Remember callers saves useful details after each call and uses them the next time the agent recognizes the same person. By default, each person is recognized by the phone number they called from. It’s off by default.

​

Voice & Understanding

Adjusts how the agent captures audio and how it prioritizes specific terms.

​

Speech Capture Quality

Sets the audio format used to process speech. Options include PCM 8000 Hz, PCM 16000 Hz (Recommended), PCM 22050 Hz, PCM 24000 Hz, PCM 44100 Hz, PCM 48000 Hz, and ulaw 8000 Hz. For most setups, leave it on PCM 16000 Hz. It’s the best balance between audio quality and processing speed. ulaw 8000 Hz is specifically designed for telephony scenarios.

​

Priority Keywords

Lets you list, separated by commas, terms the agent should prioritize when recognizing speech. Useful for brand names, product names, technical terms, acronyms, codes, or unusual surnames the model might otherwise mishear — for example SolarMax, AcmeCorp, 401k, deductible.

​

Conversation Flow

Controls the rhythm of the agent — how fast it joins in, how patient it is with slow speakers, and when it should hang up after silence.

​

Eagerness

How quickly the agent joins the conversation after the caller stops talking:

• Patient — waits longer before responding.
• Normal — balanced (default).
• Eager — responds quickly.

Normal works for most cases.

​

Spelling patience

Makes the agent more patient when the caller is slowly spelling something — names, emails, ID numbers — so it doesn’t cut them off mid-letter. Options:

• Auto — the default, recommended when the agent often collects spelled-out information.
• Off — keeps standard timing.

​

Take turn after silence

How many seconds the agent waits for the caller’s reply before speaking again (default 7). If the caller stays silent for that long, the agent re-engages them. Lower values keep the conversation moving; higher values give the caller more room to think.

​

End conversation after silence

How many seconds of silence before the call ends automatically when the caller stops responding. It helps manage costs and keeps your call logs clean. The default, -1, means the call never ends because of silence: it keeps running, and using minutes, until someone hangs up or it reaches Max conversation duration.

​

Max conversation duration

The maximum length of a single call, in minutes (default 10). A safety net to keep calls concise and prevent them from running indefinitely.

​

Background sound

Plays a continuous background sound during the conversation, mixed with the agent’s voice. Turn on Enable background sound and set:

• Sound — Office (Quiet) (default), Office (Busy), Restaurant, City, Typing, and several Elevator options. Use the play button to preview it.
• Volume — the volume of the sound relative to the agent’s voice (default 60%).
• Crossfade loop — smooths the point where the sound loops, to avoid audible pops.

​

Soft timeout

Gives the caller immediate audio feedback when the model takes longer than expected to generate a response, so the line never feels dead. Turn on Enable soft timeout and set:

• Timeout duration — how many seconds to wait for the model’s response before playing a message (default 3).
• Use LLM Generated Message — generates a contextual filler phrase with the model.
• Message — the fallback phrase used if no filler is generated in time (default “Hhmmmm…yeah.”).
• Additional filler messages — use Add message to add phrases for the 2nd, 3rd, and later soft timeouts in the same response, so the agent doesn’t repeat itself.

​

Privacy & Data

​

Record Calls

On by default. Securely stores the transcripts and audio of every call, so you can review the agent’s performance and keep a history of your customer interactions. Turn it off only if it’s not compatible with your internal privacy, storage, and compliance policies.

​

Telephony Extras

Telephony-specific settings, mostly aimed at improving outbound call performance.

​

Number Rotation

Automatically cycles through your available phone numbers. Helps prevent a single number from being flagged as spam, distributes call volume, and improves deliverability. It requires phone numbers in your account: if there are none, a warning asks you to add them first.

​

Select Phone Numbers and Calls Per Number

When Number Rotation is on and your account has numbers, you choose:

• Select Phone Numbers — which numbers will rotate. At least one is required.
• Calls Per Number — how many calls to make from one number before switching to the next.

​

State-Based Restriction

Increases answer rates by making the agent dial each contact with a phone number from their own state. Currently it supports US local numbers only — confirm your account has compatible local numbering before enabling it.