Talk to Your Agents in Real Time

Real-time voice conversations with your agents in the browser, using your own LiveKit and speech-provider accounts.

What it is

Voice is a real-time, two-way spoken conversation with an agent directly in your browser. It uses WebRTC, through LiveKit, for the audio connection. There's no telephony involved: no phone numbers, no dialing, no outbound calling.

More than one agent on a call

The agent leading a call can bring in a teammate it's allowed to delegate to, the same connection that powers text-based delegation. Once a teammate joins, either agent can hand off the conversation to the other.

Everyone shares the same audio, so agents can't talk over each other: whatever a teammate says queues on that same track and plays in order. Whoever spoke or acted most recently holds the floor; if they go quiet, there's a short window where you can keep talking to them before the conversation reverts back to the lead agent automatically. If a teammate stalls or the call runs into trouble, control always falls back to the lead rather than leaving the conversation stuck on whoever went silent.

A teammate joins as its own agent run, with its own conversation you can open and review afterward, not as a scripted response from the lead.

Bring your own providers

Voice needs two things connected before it works, both configured in Settings → Voice:

A LiveKit project: the transport that carries audio between your browser and the agent.
A speech provider for transcription and, optionally, spoken replies. Deepgram and ElevenLabs are both supported today. You can use a different vendor for each role, since coverage varies by vendor and language.

Credentials are per organization. Nothing is shared across organizations on the same install.

What you see before the agent does

Every spoken message appears in the chat as an editable transcript before it sends. If a word is misheard, you fix it before the agent ever reads it.

Voice input without voice output

An organization can enable transcription without a text-to-speech provider. That's a supported state, not a broken one: you dictate to the agent and read its text reply. The interface says so directly rather than leaving a microphone that quietly does nothing.

Language coverage

Coverage depends on which vendor is connected, not on the platform itself:

Deepgram: transcribes English, Spanish, French, and Mandarin. Speaks English, Spanish, and French only.
ElevenLabs: transcribes and speaks all four.

If no connected vendor covers a language for a given direction, the settings page states that as a fact rather than letting a session start and answer in the wrong language.

Privacy

Raw audio is never stored. Only the resulting transcript persists, the same as any other message in the conversation.

Voice Setup: connecting LiveKit and a speech provider
Delegation: the connection a lead agent uses to bring a teammate onto a call
Channels: the other ways to reach an agent
Security: how per-organization credentials stay isolated