Voice Agent Node

Configures an agent that talks to people out loud. HostWebhook does not carry the audio: it provisions the agent on ElevenLabs, Vapi or Retell, and then becomes the part that matters — the tools the agent can call, the knowledge it can look things up in, and the record of what was said.

Where the work actually happens

Worth understanding before you configure anything, because it explains why some settings live here and some live at the provider.

textProvision and proxy

Person speaks
  → provider (ElevenLabs / Vapi / Retell)     ← audio, turn-taking, latency
      → agent decides to call a tool
        → HostWebhook webhook                 ← your flow runs here
          → nodes: query a database, book a slot, send an email
        ← result goes back
  ← agent speaks the answer

The provider owns the microphone, the voice and the sub-second turn-taking — problems already solved, and not where HostWebhook adds anything. What you get here is that the agent's tools are your flows, its knowledge base is your Vector Stores, and every call leaves a transcript and a tool trace you can audit.

ℹThis node configures the agent. To make a flow run when the agent calls a tool, or when a call ends, you need the voice call trigger on the other side. This node provisions; the trigger receives.

Config

Provider and credential

ElevenLabs, Vapi or Retell, each with a credential of its own. The provider is what HostWebhook talks to when it creates and updates the agent, so switching it means provisioning somewhere new — not flipping a switch on the same agent.

LLM, system prompt, first message

LLM — the model behind the conversation, from the list the provider supports.
System prompt — who the agent is and what it must not do. On a voice agent this matters more than in chat: nobody reads a disclaimer out loud, so the rules have to be in the prompt and in the tools you do not give it.
First message — what it says before the person says anything. Empty means the agent waits, which on a phone call reads as a dead line.

Voice and languages

The voice comes from the provider's catalogue, fetched with your credential. Languages is a list, not one value — an agent that expects Spanish and English will not stumble when a caller switches mid-sentence.

Tool timeout

How long the agent waits for one of your flows to answer. This is a conversational number, not a technical one: a person on a call tolerates about a second of silence. A tool that needs eight seconds needs the agent to say something first.

Tools

The tools are what turn a voice into a colleague. Two kinds:

Built-in tools — the ones the provider ships, such as ending the call or transferring it.
Custom HTTP tools — a name, a description, a method (GET, POST, PUT, PATCH, DELETE), a URL, headers, and two kinds of parameter: path parameters that fill the {placeholders} in the URL, and query parameters appended as ?name=value.

⚠The description is the whole interface. The model decides whether to call a tool from its name and description alone — there is no schema it can inspect and no docs it can read. “Looks up an order by its number and returns its status and delivery date” gets called; “order endpoint” does not.

There is also a tool dispatch outputs surface, which is how a tool call reaches your flow rather than an external URL — the agent calls, HostWebhook dispatches into the pipeline, and the result goes back as the tool's answer.

Knowledge base

Pick one or more Vector Stores and their content is pushed to the provider as the agent's knowledge base — one provider document per source document. This is what lets the agent answer “what is your returns policy” without a tool call and without you writing the policy into the prompt.

⚠The sync is manual. Ingesting into a Vector Store does not push it to the agent — you press Sync. A store that gained twenty documents this morning is still the old corpus to the agent until someone does. Re-syncing purges what it sent before, so the agent never ends up with two copies.

RAG tuning controls how much of the retrieved material reaches the model. On a voice agent, less is usually better: every retrieved passage is prompt the model has to read before it can start speaking, and the caller hears that as a pause.

Advanced

Setting	What it decides
Conversation behavior	How the agent handles interruptions and silence.
Turn detection	When it decides you have finished speaking. Too eager and it talks over people; too patient and it feels slow.
ASR keywords	Words the recogniser should expect — product names, surnames, SKUs. The cheapest fix for “it never understands our brand name”.
Dynamic variables	Values injected per call — the caller's name, their plan — so one agent serves everyone without a prompt per customer.
LLM extras	Provider-specific model options.
Post-call evaluation	Judges the transcript after hanging up — did it resolve the issue, was the tone right.
Privacy & safety	What is recorded and retained.
Max call duration	Seconds before the call is cut. The backstop against a loop that talks to itself all night at per-minute rates.

The agent can live on your site as a microphone button in the chat widget. Widget styling is the look; access & security is who may open it; and Web widget embed hands you the snippet with the agent already wired in.

The agent slug is the public name the widget uses, so it appears in the page source — pick something you are happy for anyone to read.

Sync and webhook

Provider sync is the button that makes the configuration on this page real: it creates or updates the agent at the provider. Until you press it, everything here is a draft.

Webhook URL — where the provider sends tool invocations and the post-call payload.
Webhook secret (HMAC) — what proves a call to that URL really came from the provider. Without it the endpoint is something anyone who learns the URL can post to.

Gotchas

Editing without syncing changes nothing. The agent answering the phone is the last version that was pushed.
The provider bills the minutes, you bill the tools. A call is metered by the provider; the flows its tools trigger are events on your side. A chatty agent is expensive twice.
Voice has no undo. A tool that deletes something will be called on a mishearing sooner or later. Give the agent read tools freely and write tools deliberately.
A knowledge base is not memory. It is the same corpus for every caller. What one person said earlier in the call belongs to the conversation, not the store.