Node reference - Voice.GradientFlow

Every card on the canvas is a node. This page lists each node in the Components panel, group by group, in the order the panel shows them.

Generated from the node catalog (lib/builder/node-templates.ts). Keep in sync: when a node, label or help line changes in the catalog, update the matching section here in the same change. The docs vocabulary check fails the build when a page names a node that is not in the catalog.

How to read each entry:

What it does in one line.
Connects describes the dots on the node. “Any number” means several lines can meet at that dot.
Settings lists every field with the label you see in the editor and the help line shown under it. Fields marked required must be filled before the agent can run; the readiness pill reports them. Fields listed under More settings are hidden until you expand that section in the editor.
Needs a key names the provider whose key must be present, either from your workspace or entered in Settings.

Persona

What it does: Who the agent is: its personality, voice, and the model behind it. Every agent has exactly one Persona. Connects:

Dot	Accepts or sends	Notes
Tools (`tools_in`)	From any Tool node	Any number. Tools here are available in every Skill.
Output (`output`)	To Skill nodes	Any number.

Settings:

Setting	Help
Name (required)	What the agent calls itself on the call.
Personality and instructions (required)	Who the agent is, how it talks, and the rules it follows on every call.
Language	The language the agent listens and speaks in. Pick Multilingual to let it follow the caller. Default English.
Expected Languages	Shown when Language is Multilingual. The languages callers are most likely to use. Helps the agent recognise their speech faster.
Caller can interrupt	Off or On. When on, the agent stops talking as soon as the caller starts speaking. Default On.
Record calls	Off, On or With consent. Off keeps no audio of published calls. With consent records only after the agent has said the recording notice. Playground calls are always recorded so you can play them back. Default Off.
Recording notice	Shown when Record calls is On or With consent. Said word for word before the greeting, and the caller cannot talk over it. With consent needs this filled in.
Keep recordings for (days)	Shown when Record calls is On or With consent. Recordings older than this are deleted automatically. Leave it empty to keep them until you delete them.
Model (LLM) (required)	The AI model that decides what the agent says. Pick a provider, then a model. Recommended providers are shown first; the rest sit behind Show all providers. Under each model the list shows its relative cost ($ to $$$$) and, for models we have measured on our voice tests, the time to its first reply. Compare models opens a table with the measured first reply, test result, cost per test call and list price of every model. Models we have not measured say so.
Voice (TTS) (required)	The voice the caller hears. Pick a provider, then a voice.
Hearing (STT) (required)	How the agent turns the caller’s speech into text. Pick a provider.

More settings:

Setting	Help
Start speaking early	Lets the agent begin its reply before the full answer is ready. Only takes effect when your workspace has this feature switched on. Default Off.
Input Variables	Details you pass in when a call starts, such as the caller’s name. Write `{{name}}` inside any text to use them.
Avatar	An optional on-screen face for video calls. Leave it off for phone and voice-only agents.

Needs a key: the providers you pick under Model (LLM), Voice (TTS), Hearing (STT) and Avatar, unless your workspace covers them.

Speech-to-speech agent

Shown as Speech-to-speech agent in Components. Exported files and older flows call it Realtime Agent. What it does: An alternative to the Persona for agents built on a speech-to-speech model, where one model listens and talks without separate voice and hearing steps. Most agents use a Persona instead. An agent has either a Persona or a speech-to-speech agent, not both. Connects:

Dot	Accepts or sends	Notes
Realtime (`realtime_in`)	From one speech-to-speech model	One line. Required.
Tools (`tools_in`)	From any Tool node	Any number.
Output (`output`)	To End	One line. A Realtime Agent runs as one open conversation with no Skills: the call ends when the caller hangs up, and nothing connected after it runs, including after-the-call nodes. The readiness check reports a line to If/Else or Categorize as an invalid connection.

Settings:

Setting	Help
Name (required)	What the agent calls itself on the call.
Personality and instructions (required)	Who the agent is, how it talks, and the rules it follows on every call.
Language	The language the agent listens and speaks in. Pick Multilingual to let it follow the caller. Default English.
Expected Languages	Shown when Language is Multilingual. The languages callers are most likely to use.

More settings:

Setting	Help
Caller can interrupt	Always On. Speech-to-speech models always let the caller interrupt.
Input Variables	Details you pass in when a call starts, such as the caller’s name. Write `{{name}}` inside any text to use them.

Needs a key: the provider of the connected speech-to-speech model.

Speech-to-speech

Speech-to-speech models for a speech-to-speech agent. Each has one output dot that connects to the agent’s Realtime dot. Components shows them as OpenAI speech-to-speech, Gemini speech-to-speech, Azure OpenAI speech-to-speech and xAI speech-to-speech; exported files and older flows use the names below.

OpenAI Realtime

What it does: Speech-to-speech via the OpenAI Realtime API. Settings: Model (default GPT Realtime 1.5), Voice (default Alloy), Temperature (0.6 to 1.2, default 0.8). Needs a key: OpenAI.

Gemini Realtime

What it does: Speech-to-speech via the Gemini Live API. Settings: Model (default Gemini 3.1 Flash Live), Voice (default Puck), Temperature (0 to 2, default 0.8). More settings: Affective Dialog (2.5 only), default Off. Needs a key: Gemini.

Azure OpenAI Realtime

What it does: Speech-to-speech via Azure OpenAI Realtime. Settings: Deployment (your Azure deployment name), Voice (default Alloy), Temperature (0.6 to 1.2, default 0.8). More settings: Endpoint. Needs a key: Azure OpenAI.

xAI Realtime

What it does: Speech-to-speech via the xAI Grok Voice Agent API. Settings: Voice (Ara, Clio or Sage; default Clio). Needs a key: xAI.

Conversation

Skill

What it does: One part of the conversation: what the agent does in this phase. The agent follows the Instruction until the phase is done, then moves to the next connected step. A Skill never ends the call by itself; every path ends with an End node. Connects:

Dot	Accepts or sends	Notes
From (`persona_in`)	From the Persona, another Skill, If/Else or Categorize	One line.
Tools (`tools_in`)	From any Tool node	Any number. Tools here are available only in this Skill.
Output (`output`)	To another Skill, If/Else, Categorize or End	Any number.

Settings:

Setting	Help
Instruction (required)	What the agent should do and say during this part of the call.
Skill Name	A short name for this step. Shown on the canvas and in call transcripts.
Greeting	The opening line when the call starts. Only the first Skill after the Persona uses it.
Say exactly this	Speak the greeting word for word instead of letting the agent rephrase it. Default Off.
When To Use	How the agent decides to use this Skill.

More settings:

Setting	Help
What This Skill Is Good At	A one-line summary of what this step handles. Helps the agent pick the right step.
Completion Criteria	How the agent knows this step is done and it can move on.
Handoff Notes	Details collected here that the next step should remember and not ask again.
Time limit	Seconds before the agent moves to the next connected step. 0 means no limit. If nothing is connected, the call ends.
Turn limit	Caller replies before the agent moves to the next connected step. 0 means no limit. If nothing is connected, the call ends.
Can Transfer To Human	Lets the agent offer a phone transfer from this step. Needs a SIP Transfer tool with a number connected here; a SIP Transfer on the Persona already allows it everywhere. Human Handoff does not use this setting. Default Off.

Needs a key: none.

If/Else

What it does: Checks a condition and sends the call down the True or False path. Connects:

Dot	Accepts or sends	Notes
Previous Step (`in`)	From a Skill	One line.
Different model (`llm_in`)	From an LLM (Override) node	Optional. Uses the Persona’s model otherwise.
True (`true`)	To a Skill or End	One line.
False (`false`)	To a Skill or End	One line.

Settings:

Setting	Help
Condition	Written in plain words. When it holds, the call follows the True path, otherwise the False path.

Needs a key: none.

Categorize

What it does: Sorts the caller’s request into a category and sends the call down that path. Connects:

Dot	Accepts or sends	Notes
Previous Step (`in`)	From a Skill	One line.
Different model (`llm_in`)	From an LLM (Override) node	Optional. Uses the Persona’s model otherwise.
One output per category	To a Skill or End	One line each. A new output appears for each category you add. A caller sorted into a category with no line is hung up, so the readiness check warns until every output is connected.

Settings:

Setting	Help
How to sort	Tell the agent what to look at when choosing a category.
Categories	Each category adds its own output to the step. Connect each output to whatever should handle that category. Removing a category removes its connection; the ones below it keep theirs. At least two. Default Sales, Support.

Needs a key: none.

End

What it does: Ends the conversation and hangs up the call. A call only ends when it reaches an End node. Connects:

Dot	Accepts or sends	Notes
In (`in`)	From a Skill, If/Else or Categorize	One line.
Output (`output`)	To after-the-call nodes: Webhook, Structured Output, Summarization, CRM sync (after call), Helpdesk	Any number.

Settings:

Setting	Help
Goodbye message	What the agent says before hanging up. Leave blank for a short default goodbye.
Say exactly this	Speak the goodbye word for word instead of letting the agent rephrase it. Default Off.

Needs a key: none.

Tools

Tools give the agent something it can do during the call. Every tool has one output dot (output) that connects to the Persona (all Skills) or to one Skill. Tools that call an outside service stop after 10 seconds by default: the agent tells the caller it could not finish, and a tool that sends or changes something is not repeated, since it may still have gone through. SIP Transfer has no default limit, because it waits for a person to answer.

MCP

What it does: Connects to an MCP server to give the agent its tools. Settings:

Setting	Help
Server URL (required)	The address of the MCP server. Without it the agent gets no tools from this node.
Auth Type	None, Bearer Token or API Key. How the server expects to be signed in to. Leave as None for open servers.
Auth Value	Shown when Auth Type is not None. The token or key the server gave you.
Tools	Which of the server’s tools the agent may use. Leave empty to allow all of them.

More settings:

Setting	Help
Time limit (seconds)	How long the agent waits for this tool before telling the caller it could not finish. 1 to 120, default 10.
While running, say	Said once if the tool is still working after a second and a half. Leave blank to stay quiet. Default blank.

Needs a key: none from Voice.GradientFlow. See MCP servers.

Exa Web Search

What it does: Lets the agent search the web for up-to-date answers. Settings:

Setting	Help
Search Type	Auto, Neural or Keyword. Auto picks the best search style for each question. Default Auto.
Number of Results	How many pages the agent reads for each search. More is slower. 1 to 10, default 5.
Include Page Contents	Read the page text, not just titles, so answers are more accurate. Default On.

More settings:

Setting	Help
Time limit (seconds)	How long the agent waits for this tool before telling the caller it could not finish. 1 to 120, default 10.
While running, say	Said once if the tool is still working after a second and a half. Leave blank to stay quiet. Default blank.

Needs a key: Exa. See Web search.

Pinecone

What it does: Lets the agent answer from your own documents stored in Pinecone. Settings:

Setting	Help
Index Name (required)	The name of the Pinecone index that holds your documents.
Namespace	A section inside the index, if you use them. Leave blank for the default.
Top K Results	How many matching passages the agent reads per question. 1 to 10, default 3.

More settings:

Setting	Help
Embedding Model	Must match the model your index was built with. Default Pinecone multilingual-e5-large.
Time limit (seconds)	How long the agent waits for this tool before telling the caller it could not finish. 1 to 120, default 10.
While running, say	Said once if the tool is still working after a second and a half. Leave blank to stay quiet. Default blank.

Needs a key: Pinecone.

SIP Transfer

What it does: Dials a phone number and connects the caller to a person. Settings:

Setting	Help
Phone Numbers (required)	The numbers the agent can transfer to. Without one the transfer is switched off.
Outbound trunk	The phone line transfers dial out on, from the outbound trunks under Phone. Automatic uses your workspace’s only outbound trunk. Default Automatic.
Transfer type	How the caller reaches the person: Join the call (the person joins and the agent goes quiet; the whole call stays recorded), Brief them first (the agent calls the person, says who is waiting and why, then connects the caller), or Hand the call over (your carrier moves the call; recording stops there). Default Join the call.
If nobody answers	Brief them first only. Go back to the caller, or Leave a voicemail first. Either way the agent goes back to the caller and keeps helping. Default Go back to the caller.
What to tell the person (More settings)	Brief them first only. Added to the summary the agent gives before connecting. Optional.
Ring for (seconds) (More settings)	Brief them first only. How long the person’s phone rings before the agent goes back to the caller. 10 to 120, default 30.
Play Dial Tone	Join the call and Hand the call over only. Let the caller hear ringing while the transfer connects. Brief them first plays hold music instead. Default Off.

Needs: an outbound trunk under Phone in the dashboard. See SIP transfer.

Simulated action

What it does: Simulated action for testing. Replace it with a real tool before going live. Settings:

Setting	Help
Tool name (required)	A short name for the pretend action, in lowercase with underscores.
What this action does	Tells the agent when to use this action. Nothing really happens; it always succeeds.
Parameters	The details the agent must collect before using the action, such as a date or name.
Success message	What the agent is told after the pretend action succeeds.

Needs a key: none.

Human Handoff

What it does: Lets a person on your team join the call and take over. The call is flagged in Dashboard, Handoffs and the agent goes quiet while the person talks. Settings:

Setting	Help
What the agent says while handing off	The agent will phrase this in its own words.
Resume After Human Leaves	When the person leaves the call, the agent picks the conversation back up. Default On.

Needs a key: none. Someone must be watching the Dashboard.

Calendar

What it does: Lets the agent check availability and book appointments in Cal.com. Settings:

Setting	Help
Event Type ID (required)	The number of the Cal.com event type to book. Find it in the event type’s web address.
Timezone	Times are spoken to the caller in this timezone. Default UTC.
Confirm Before Booking	The agent reads the slot back and asks the caller to confirm before booking it. Default On.

More settings:

Setting	Help
Max Slots Offered Per Day	How many free times the agent offers for one day. Default 6.
API Base URL	Only needed if you run your own Cal.com server.
Time limit (seconds)	How long the agent waits for this tool before telling the caller it could not finish. 1 to 120, default 10.
While running, say	Said once if the tool is still working after a second and a half. Leave blank to stay quiet. Default blank.

Needs a key: Cal.com. See Calendar booking.

CRM lookup (during call)

What it does: Lets the agent look up and update HubSpot contacts while talking. Settings:

Setting	Help
Allow Writes	Let the agent create or update contacts, not just look them up. Default On.

More settings:

Setting	Help
Time limit (seconds)	How long the agent waits for this tool before telling the caller it could not finish. 1 to 120, default 10.
While running, say	Said once if the tool is still working after a second and a half. Leave blank to stay quiet. Default blank.

Needs a key: HubSpot. See CRM.

What it does: Sends WhatsApp messages through Twilio during or after a conversation. Settings:

Setting	Help
Fixed Recipient	Always send to this number. Leave blank to let the agent ask the caller for one.
Confirm Before Sending	The agent reads the message back and asks the caller to confirm before sending. Default On.

More settings:

Setting	Help
Time limit (seconds)	How long the agent waits for this tool before telling the caller it could not finish. 1 to 120, default 10.
While running, say	Said once if the tool is still working after a second and a half. Leave blank to stay quiet. Default blank.

Needs a key: Twilio.

SMS

What it does: Sends text messages through Twilio during or after a conversation. Settings:

Setting	Help
Fixed Recipient	Always send to this number. Leave blank to let the agent ask the caller for one.
Confirm Before Sending	The agent reads the message back and asks the caller to confirm before sending. Default On.

More settings:

Setting	Help
Time limit (seconds)	How long the agent waits for this tool before telling the caller it could not finish. 1 to 120, default 10.
While running, say	Said once if the tool is still working after a second and a half. Leave blank to stay quiet. Default blank.

Needs a key: Twilio.

Email

What it does: Sends emails from your mail server during or after a conversation. Settings:

Setting	Help
Fixed Recipient	Always send to this address. Leave blank to let the agent ask the caller for one.
Confirm Before Sending	The agent reads the message back and asks the caller to confirm before sending. Default On.

More settings:

Setting	Help
Time limit (seconds)	How long the agent waits for this tool before telling the caller it could not finish. 1 to 120, default 10.
While running, say	Said once if the tool is still working after a second and a half. Leave blank to stay quiet. Default blank.

Needs a key: a mailbox configured by your workspace operator.

Post Processing

These nodes run after the call ends. Connect them after End. Structured Output and Summarization can also feed the others.

Webhook

What it does: Sends the call data to a web address when the call ends. Connects: In (in) from End, Structured Output or Summarization. Any number of lines. Settings:

Setting	Help
Send to	Any web address, Zapier or Make. Zapier and Make only change the field label and the help text; the request is the same. Default Any web address.
URL (required)	Where to send the call data. Your system receives the transcript and any extracted fields, signed with the secret in Settings → Webhooks. For Zapier, the Catch Hook address from your Zap; for Make, the Custom Webhook address from your scenario.

Needs a key: none. See Webhooks.

Structured Output

What it does: Pulls specific details out of the conversation after the call. Connects:

Dot	Accepts or sends	Notes
End (`end`)	From End	Any number.
Different model (`llm_in`)	From an LLM (Override) node	Optional. Uses the Persona’s model otherwise.
Output (`output`)	To Webhook, CRM sync (after call) or Helpdesk	Any number.

Settings:

Setting	Help
Name (required)	The label these details are filed under when they are sent on.
Extraction Instruction	Tell the agent what to look for in the transcript.
Output Schema (required)	The fields to fill in, such as name, email, or order number.

Needs a key: none. See Structured output.

Summarization

What it does: Writes a short summary of the call after it ends. Connects:

Dot	Accepts or sends	Notes
End (`end`)	From End	Any number.
Different model (`llm_in`)	From an LLM (Override) node	Optional. Uses the Persona’s model otherwise.
Output (`output`)	To Webhook, CRM sync (after call) or Helpdesk	Any number.

Settings:

Setting	Help
Name	The label the summary is filed under when it is sent on.
Summary Style	Bullet points, Paragraph, or Brief (1 to 2 sentences). How the summary is written. Default Bullet points.
Instructions	Anything the summary should focus on or leave out.

Needs a key: none.

CRM sync (after call)

What it does: Saves the call to the caller’s HubSpot contact after it ends. Connects: In (in) from End, Structured Output or Summarization. Any number of lines. Connect a Structured Output that extracts the caller’s email so the contact can be matched. Settings:

Setting	Help
Email Field	The name of the extracted field or input variable that holds the caller’s email. Default email.
Create Contact If Missing	Add a new HubSpot contact when the caller is not found. Default On.
Log Conversation Note	Attach the transcript to the contact as a note. Default On.

Needs a key: HubSpot. See CRM.

Helpdesk

What it does: Opens a support ticket after the call ends. Connects: In (in) from End, Structured Output or Summarization. Any number of lines. Settings:

Setting	Help
Where to file tickets	HubSpot Tickets or This dashboard. HubSpot needs a HubSpot key. The dashboard option needs no account. Default HubSpot Tickets.
Subject	The title of every ticket this agent opens. Default “Voice call follow-up”.
Priority	Low, Medium or High. The priority set on every ticket this agent opens. Default Medium.
Email Field	The name of the extracted field or input variable that holds the requester’s email. Default email.

More settings:

Setting	Help
HubSpot Pipeline ID	Only if your HubSpot uses more than one ticket pipeline.
HubSpot Pipeline Stage ID	The stage new tickets start in.

Needs a key: HubSpot for HubSpot Tickets. None for This dashboard. See Helpdesk tickets.

Advanced

LLM (Override)

What it does: Only for If/Else and Categorize decisions when they should use a different model than the Persona. Most agents do not need this. It can also feed Structured Output and Summarization. Connects: Output (output) to the Different model dot on If/Else, Categorize, Structured Output or Summarization. Any number of lines. Settings:

Setting	Help
Model	The model used for these decisions instead of the Persona’s model. Default gpt-5.4-mini.

More settings:

Setting	Help
Temperature	How much the wording varies from call to call. Lower is more predictable. Default 0.7.
Max Tokens	Caps how long one reply can be. Leave blank for the default.

Needs a key: OpenAI.

​Persona

​Persona

​Speech-to-speech agent

​Speech-to-speech

​OpenAI Realtime

​Gemini Realtime

​Azure OpenAI Realtime

​xAI Realtime

​Conversation

​Skill

​If/Else

​Categorize

​End

​Tools

​MCP

​Exa Web Search

​Pinecone

​SIP Transfer

​Simulated action

​Human Handoff

​Calendar

​CRM lookup (during call)

​WhatsApp

​SMS

​Email

​Post Processing

​Webhook

​Structured Output

​Summarization

​CRM sync (after call)

​Helpdesk

​Advanced

​LLM (Override)

Persona

Persona

Speech-to-speech agent

Speech-to-speech

OpenAI Realtime

Gemini Realtime

Azure OpenAI Realtime

xAI Realtime

Conversation

Skill

If/Else

Categorize

End

Tools

MCP

Exa Web Search

Pinecone

SIP Transfer

Simulated action

Human Handoff

Calendar

CRM lookup (during call)

WhatsApp

SMS

Email

Post Processing

Webhook

Structured Output

Summarization

CRM sync (after call)

Helpdesk

Advanced

LLM (Override)