Configure assistant

Open the builder

Go to Assistants and open an agent. The left sidebar lists every settings section. Changes stay unsaved until you click Save. Use Talk / Phone test after saving to verify behavior.

Assistant

This is the identity and brain of the agent.

• AI Language Model — On Voxovo 1 this shows the managed stack (Speed / Smart / Ultra) and the underlying model. On Custom you pick any catalog LLM.
• Voice & Language — Who the agent sounds like and which language pack to lock. Changing language does not overwrite a voice you already chose.
• Agent name / First message — Display name and the opening line (or wait for the caller).
• System prompt — Core instructions. Use Generate Prompt to analyze your settings + draft and enhance it with AI.
• Responsibilities — Roles like receptionist, booking, support. These are written into the live prompt so the agent knows what it must handle.
• Qualifying questions — Questions the agent should ask to qualify the caller.

Voice Settings

Controls the speech pipeline: how audio is heard, reasoned, and spoken.

• Voxovo 1 pipeline — Managed Speed / Smart / Ultra stacks. You mainly pick language and voice; models stay locked to that tier.
• Custom pipeline — Choose your own STT (speech-to-text), LLM, and TTS (text-to-speech) providers and models — same card style used for ambience scenes.
• STT (transcriber) — Turns caller audio into text. Phone lines prefer telephony-tuned models (e.g. Deepgram nova-2-phonecall).
• TTS (voice) — Turns agent text into speech. Preview with the play button before you save.

Call Settings

Behavior for live phone / browser calls.

• End call — Auto hangup phrases, max duration, silence timeout, and whether to record calls.
• Human transfer — Warm/cold transfer to a human number, hold message, and supervisor desk for live listen/whisper.
• Advanced call controls — Voicemail detection (AMD), DTMF keypad input, language detection, ring timeout / attempts, caller ID label.

Background ambience

Adds a soft room tone under the agent voice so calls feel natural (office, cafe, rain, etc.). These are procedural beds — not stock music clips.

• Off — quiet line — No ambience; agent voice only.
• Scene cards — Same selection style as Custom pipeline LLM/TTS cards. Click a scene to select it.
• Hear — Plays a short sample of that scene at your current mix level so you can audition before Save.
• Ambience level — How loud the bed sits under speech. Start near 20–25%; higher is more present but can compete with the voice on phone lines.

Remember to Save after choosing a scene. Ambience only plays on live calls after the agent is saved and the call pipeline loads those settings.

Advanced Settings

Fine-tunes turn-taking, expressiveness, and presence while tools run.

• Conversation flow — Allow interruptions (barge-in), skip turns, end calls, interruptions during greeting, auto noise reduction.
• Voice performance — Expressiveness and how quickly fragments flush to TTS.
• Advanced AI — Extra LLM/temperature style knobs when exposed.
• Presence Buffer — Short keep-alive phrases (“one sec…”, “still looking…”) while tools or the model think, so callers do not hear dead air.

Voice Activity Detection (VAD)

VAD decides when the caller starts and stops speaking. It listens for energy / speech patterns in the audio stream and tells the pipeline “user is talking” vs “user finished.”

• Why it matters — Too sensitive → agent cuts off mid-sentence or reacts to noise. Too slow → long awkward pauses after the caller finishes.
• Speech threshold (sensitivity) — How hard it is to trigger “speech started.” Higher = ignores more noise / quieter speech.
• Silence timeout — How long quiet must last before the turn ends and the agent replies.
• Minimum speech duration — Ignores blips shorter than this (coughs, clicks).
• Prefix padding / max speech — How much audio to keep before speech start, and a hard cap on continuous speech length.

Defaults work for most phone lines. Only tune VAD if you hear cutoffs, late replies, or false barge-ins.

Squad

Connect multiple specialist assistants to this one. Add members with a When to transfer rule, enable squad handoffs, and mirror those rules in the system prompt (e.g. intake agent → billing specialist). Mid-call handoff keeps shared memory / transcript — different from phone-number warm transfer in Call Settings.

Dynamic Variables

Named slots like {{caller_name}} you can inject into the first message or prompt. Useful for personalized greetings from CRM or campaign data.

Tools / Knowledge / MCP

The Tools nav opens with three buttons (same pattern as Test): Tools (default), Knowledge, and MCP.

• Tools — Attach built-in or custom tools (calendar, sheets, webhooks). The tool description explains WHAT it does; your prompt should say WHEN to use it.
• Knowledge — Docs, facts, and FAQs the agent can search for grounded answers.
• MCP — Model Context Protocol servers / tools for deeper integrations.

Automation

Link a step-based call flow from Automations (trigger → voice agent → actions). Default editor is an ordered step list; Canvas remains available for advanced graphs. Use Squad for mid-call specialist handoffs.

Agent Memory

Optional notes the agent carries across calls (learned preferences, recurring facts). Keep short and factual.

Save and publish

Click Save often. Toggle Publish when the agent is ready for production routing. Live calls recompile settings into the system prompt so responsibilities, knowledge, and call settings stay in sync with what you configured.