Voice

Voice lets end users talk to your agent out loud and hear it talk back, in real time, instead of typing. It's a natural spoken conversation layered on top of the same chat — the agent keeps the same knowledge, tools, and personality; it just gains a microphone and a voice.

What you can do with it

• Hold a live spoken conversation with the agent: you speak, it understands, and it replies with synthesized speech.
• Switch fluidly between voice and text in the same conversation — the dialog history carries over both ways, so you can start typing and finish by talking, or the reverse.
• Spoken turns are transcribed and saved into the conversation history like any other message, and the conversation is auto-titled from the first thing said.
• Tools still work during a voice call — for example, the agent can generate an image mid-conversation, and the result shows up in the chat.
• Interrupt the agent by talking over it. Playback stops right away — even while the agent is busy running a tool — and the agent keeps only the part of its reply you actually heard, so the conversation picks up from there rather than from words that never reached you.
• Choose the agent's voice. When building the agent you can pick from the platform's available realtime voices and preview each one — Donkit plays a short sample spoken in that exact voice, and can preview it in a chosen language, before you commit. End users of a published agent may also see a voice picker driven by the same catalog.

How you turn it on

1. In the agent's settings, enable Voice. Voice is purely additive — turning it on doesn't change anything about the agent's existing text chat.
2. Choose a voice, using the preview button to audition options.
3. Publish. End users then get a microphone in the agent's chat and can start a spoken conversation.

Voice is also available in the main chat — the same conversation you land in after sign-up — with nothing to turn on: it comes enabled by default. In the main chat, voice sees the same roster of your other agents as the typed chat does, and can hand off a request to one of them mid-conversation exactly like a typed turn can. It is limited by the same realtime-voice-model availability as any other agent's voice.

Voice requires the deployment to have realtime voice models configured — on installs without them (mainly self-hosted), the microphone simply doesn't appear.

Voice also works in the phone and Mac apps, not just on the web. The app shows a live transcript alongside the tool steps the agent ran, and you can leave a spoken conversation for the text chat at any point and find the whole exchange already there.

On iPhone, iPad and Mac, changing an agent's Voice settings updates its open chat after the change is saved. Returning to the chat or bringing the app back to the foreground also refreshes voice availability, without starting a new chat.