Glossary

Terms used across these guides, in plain language.

Voice pipeline

ASR — Automatic Speech Recognition, also called speech-to-text. The model that turns the caller's audio into the text your agent reads. The agent never hears audio; it only ever reads what ASR produced, including any mistakes. See How a voice agent works.

TTS — Text-to-Speech. The model that turns the agent's reply into audio. Which provider you pick affects latency and cost as much as it affects how the voice sounds. See Choosing a voice.

LLM — Large Language Model. The model that reads the conversation and decides what to say or which tool to call. Set per agent on Agent Overview.

VAD — Voice Activity Detection. The continuous judgement of whether incoming audio is speech or just noise, and — the part that matters most — the moment the caller has stopped talking. Most of a call's perceived responsiveness comes from this one decision.

Barge-in — the caller starting to speak while the agent is mid-sentence. When it fires the agent stops immediately and discards the rest of what it was going to say. Distinguishing a real interruption from a listening noise like "mm-hmm" is one of the hard problems in voice.

Turn — one exchange: the caller speaks, the agent replies. A typical call is five to forty of them. Conversations show a turn count per call.

Turn-taking — deciding whose turn it is to speak. See How a voice agent works.

Echo cancellation — stops the agent hearing its own voice through the caller's speaker and mistaking it for the caller talking.

Latency — the silence between the caller finishing and the agent starting. It accumulates across every pipeline stage.

Platform objects

Workspace — the container everything belongs to: agents, numbers, contacts, campaigns, integrations. Nothing is shared between workspaces. See Workspaces.

Agent — the thing that talks: a prompt, a voice, a model, knowledge and tools. Internally a VoiceWidget, which is why API payloads say widget_id.

System prompt — the instructions defining who the agent is, what it does and what it refuses. The single highest-leverage field. See Writing the system prompt.

First message — the agent's opening line. Leave it empty and the agent waits for the caller to speak first.

Knowledge base — documents the agent searches during a call, as a tool call. Not loaded into the prompt. Attached per agent. See Knowledge base.

Tool — an action the agent can take beyond speaking: end the call, transfer it, look up a contact, book a slot. See Tools.

Campaign — a scheduled batch of outbound calls. Only campaigns set to active are ever dialed. See Outbound campaigns.

Contact — a person the agent can call, deduplicated per workspace by phone number, carrying consent state.

Conversation — the record of a call: transcript, tool calls, timing, recording, how it ended.

Telephony

SIP — Session Initiation Protocol, the standard for setting up voice calls over IP. Connect an existing SIP trunk instead of buying numbers through Twilio or Telnyx.

SIP trunk — the connection between iCallAgent and a carrier that carries your calls.

BYOC — Bring Your Own Carrier. Routing outbound calls through your own carrier contract rather than the platform's. Configured as Outbound Trunk (BYOC) under Integrations.

Extension — a short dialable number that reaches one agent directly over a SIP trunk. Assigned automatically to every agent.

Caller ID — the number shown to the person you're calling. Set from the phone number attached to the campaign.

Inbound / outbound — a call to your number, or one your agent places.

Hangup cause — the reason a call ended, recorded on every finished call. Useful for telling a busy line apart from a rejected one.

Compliance

Consent — whether a contact has agreed to be called: granted, implied, unknown or revoked. Checked before every outbound call.

Do-not-call (DNC) — a suppression list of numbers that must never be dialed. Maintained per workspace under Settings → Compliance.

Blacklist — a per-contact block, separate from the DNC list.

Compliance gate — the check that runs before any outbound call, covering blacklist, DNC and consent. A blocked contact is never dialed, and the API rejects the attempt with a reason.

AI disclosure — telling the person they're speaking to an AI. Required in several jurisdictions; campaigns carry a disclosure script field for it.

Developer

Widget — the browser call button you can embed on your own site, as an alternative to a phone number.

Allowed domain — the one origin permitted to embed an agent's widget. Set automatically to the dashboard when the agent is created.

MCP — Model Context Protocol, the mechanism the agent runtime uses to reach its tools during a call. See MCP.

Webhook — an HTTP callback iCallAgent sends you when something happens, currently call.completed. See Webhooks.

Personal API key — a bearer credential for your own workspace, created under Settings → API Keys.

OAuth2 — the flow a third-party app uses to act on another user's workspace with their consent. See Authentication.