Voice

Voice in Deska is not a separate dictation box bolted on the side. It is the front door to Dex. A compact center sits at the bottom of the canvas where you can either type a question or hold to talk, and it carries you through the whole exchange: you speak, Dex listens, thinks, and speaks its answer back, all in one place.

It moves through a few clear states so you always know what is happening:

• Listening: a live waveform shows it is hearing you. Press Esc to cancel or Enter to send.
• Thinking: your words have been transcribed and handed to the model.
• Speaking: the reply streams in, with small chips noting any actions Dex took on the canvas and a link to open the full thread.

A finished reply ends with Copy, which puts the whole answer on the clipboard in one click, as markdown with its code blocks intact, and with who answered and the model beside it. Mentions of panels and tasks are copied as their names. Earlier replies on the card and every reply in the expanded conversation end the same way, with the mark of the agent that answered.

The expanded conversation has two sizes. Make larger in its header widens it and gives the thread most of the window’s height; Deska remembers the size you left it at. The steps Dex took fold into one line, such as 3 steps, once the answer arrives; open it to see each step with an icon for what it touched. A long message of yours shows its first six lines with Show more. When Dex asks you to confirm something or to choose while the conversation is open, the question takes the place of the message box, so what led to it stays readable above. The chip under the message box shows the agent that answers, with its mark and its model.

You can also dictate while composing a typed message. Tap the mic, speak, and the transcription drops into the text field for you to edit before sending.

The fastest way to reach Dex by voice is to hold a key. Press and hold fn on macOS, speak, and release to send. On Windows, hold CtrlWin the same way, and on Linux the same chord is labeled CtrlSuper; the talk key is still in beta on both. On Linux it needs an X11 or XWayland session, because a native Wayland session hides global keys and Settings will say the listener is unavailable. The voice center takes over while the key is down and hands your words straight to the assistant.

The talk key is context aware. Hold it while a text field has focus and it dictates into that field, typing what you say at the cursor; hold it anywhere else and it talks to Dex, who does something and gives you an answer back. A quick tap cancels, and Esc cancels a dictation in progress, so nothing is sent.

It works from any app, and Deska stays in the background. Hold the key while you are in your browser, your editor or a chat app and Deska never comes to the front. A small pill appears at the bottom of the screen instead, showing that Dex is listening, working or answering, with its own buttons to cancel, finish, answer a confirmation or end a hands-free conversation. While a hands-free conversation is muted the pill reads Microphone muted; muting and unmuting happen in the capsule in Deska. While Dex works it is one short line with Dex’s face; Minimize folds it to the face alone, and it opens again when the answer arrives or Dex needs you. Minimizing never stops the work: to stop a turn, open Deska. While Dex answers out loud, the pill shows the sentence it is saying, never one it has not reached yet. Click the pill itself to open Deska. Drag it to put it anywhere on the screen; it stays there next time, and dropping it near the bottom center (or Settings → Keybindings → Push to talk → Reset position) sends it home. On macOS, holding the key inside another app’s text field types what you say there (Deska pastes it and puts your clipboard back); anywhere else it talks to Dex, who answers out loud. On Windows and Linux the key always types into the app you are in, and a double tap or the voice shortcut reaches Dex. The voice shortcut (CmdShiftV by default) no longer brings Deska forward either.

The first time you use voice, Deska opens a short setup instead: it asks for the microphone, for Accessibility on a Mac (so the key works in every app and Dex can type for you), points you at the globe key setting, and ends with a field to try a dictation in. It offers itself on the first voice use of a few sessions until you finish it, and Settings → Keybindings → Push to talk → Run setup opens it any time.

Some Macs have no Apple fn key, and third-party external keyboards never send it. For those, open Settings → Keybindings → Push to talk and pick ControlOption as the hold key instead.

Hold to talk is one exchange. For a longer back and forth, double-tap the talk key instead and Dex opens a continuous session with nothing held down. The microphone stays open until you double-tap the key again, say “stop Deska”, or press End. To pause it without ending it, press Mute microphone; see Mute for a moment below.

It runs in the background. The session lives in the Dex capsule at the bottom of the canvas, in the same spot and at the same size, so nothing covers your work. Keep clicking through panels and typing while you talk: clicks and Esc belong to whatever you are doing and never end the session. The capsule shows what you are saying, then what Dex is doing, then the sentence he is speaking. Click it to open the whole conversation.

It is not turn-based. Dex is listening while he is talking, so you can cut him off mid-sentence just by speaking. He stops, listens, and answers the new thing. He does not finish the sentence you interrupted and does not pick it back up. The same is true while he is thinking: start talking and the turn is abandoned for whatever you just said.

Your words appear in the capsule as you say them, rather than after you stop.

Pause as you need to. Dex takes your request once you have finished talking, about a second and a half after your last word, so stopping to think in the middle of a sentence does not send half of it. Only a voice counts as talking: typing, a fan or traffic in the background does not keep him waiting for you. Another voice about as loud as yours, like a television right next to you, still does.

Speak as soon as hands-free starts

Dex listens as soon as you activate the session: no wake name needed. After 40 seconds of silence, the capsule shows “Say Dex first” and you need to say “Dex” to resume the conversation. The timer pauses while you speak or Dex processes or answers. Start speaking before the deadline and the whole phrase is accepted, even if its transcript arrives later. There is no spoken warning or countdown. Outside that window, a phrase without his name is shown as heard and let go, without becoming a request.

A correction said within five seconds of the last one joins it rather than starting a second request: “Dex, check the agents… the ones on the old project” is one question.

While you do something else

A session survives switching workspace and does not time out, however long you stay quiet. While nobody is talking, nothing is uploaded or billed, even in a noisy room. While it is open, Dex will speak up when an agent is blocked and waiting on you, and he waits for a gap rather than talking over you. Nothing else interrupts out loud; the rest stays in the Inbox. After he speaks, the 40-second window is open, so you can just answer.

Mute for a moment

For a phone call or someone walking in, press Mute microphone, the microphone button to the left of End in the capsule. Deska lets go of the microphone, so your system’s mic indicator turns off, and the capsule reads Microphone muted. What you said before pressing it finishes transcribing and reaches Dex once, even if you mute immediately after your last word or halfway through a sentence. Dex keeps working and speaking; muting does not cancel your request. The capsule shows the transcription or work still in progress. While muted, Dex hears nothing new, not even “stop Deska”; End or a double tap still ends the session. A muted session does not time out. If the saved speech cannot finish transcribing, Deska shows an error and asks you to start another conversation and repeat the request; an unfinished transcript is never used as an instruction.

Press Unmute to pick up where you left off. The microphone takes about a second to open again, and the 40-second window opens with it, so your next sentence needs no “Dex” first.

Actions without a confirmation

When you ask Dex to close panels, delete a note, revert an agent’s work, merge its branch, remove a worktree or remove a community panel, he does it: asking was the approval. To be asked first instead, turn off Settings → Dex → Act without asking. With it off, a yes or no card appears before each of those, and during a hands-free session you can answer it out loud: say “yes” or “no” (“sí”, “dale”, “cancela” work too), no name needed. Closing panels only asks when one of them has unsaved changes.

Speakers, headphones and the echo

Interrupting by voice needs your microphone to hear you and not Dex. Most setups cancel his voice on their own. When yours does not, the session notices within two sentences, says so in the capsule, and lets him finish what he is saying rather than cutting him off on his own words. Plug in headphones and interrupting works again.

If the connection drops, the capsule says it is reconnecting and the microphone stays open. The session comes back on its own.

It is the same key as push to talk, so whichever hold key you picked under Settings → Keybindings → Push to talk is also the one you double-tap.

Because voice flows into Dex, speaking is not limited to dictating text. You can ask for work to be done. Say “run the tests” or “what is failing in the build terminal?” and Dex uses what is open on your canvas to answer, or hands the task to a coding agent and reports back. It is the same assistant you reach with Cmd/CtrlL, just driven by your voice.

Dex answers with shows each agent’s logo and an icon for its computer or Cloud host. Pick an agent, then a model when available. The closed selector names the remote computer too, so switching between two hosts using the same agent is visible.

Dex can hand work to any agent set up on this computer (Claude Code, Codex, OpenCode, Cursor, Antigravity, Pi, Hermes, Mistral Vibe, and Kimi Code), and which one takes it is yours to choose, in one place: Manage agents in the audio popover opens Settings, Agents, where New tasks use decides the agent for the tasks Dex hands off and each agent’s default model applies. If you had picked an agent or model in the popover’s former Coding agent row, Deska kept it as Agents opened in a terminal on that page.

To capture text rather than ask a question, use the dictation shortcut: CmdShiftV on macOS, or CtrlAltV on Windows and Linux. Because it is registered globally, it works even when Deska is in the background, so you can capture a thought without switching windows first. Inside Deska there is an even shorter path: hold the talk key while a text field has focus and it types what you say straight at the cursor.

A task’s own box has a mic beside its send button: a new task, a task you are following up and a Dex conversation on the board. Tap it, speak, and tap Listening to stop: the mic reads Transcribing with a spinner while your words come back, then they land at the cursor for you to edit. Send stays disabled until they arrive. To send them as they are, press the send button instead while the mic listens. It reads Stop & send (Stop & start on a new task, Stop & queue while a turn runs, Stop & answer for an agent’s question), and one press, or Enter, stops the recording, waits for the words and sends them together with what you typed. During this wait, only the send button shows Transcribing; the mic disappears and send actions are disabled. Steer now does the same into the running turn. Press Esc while it waits to keep the words in the box without sending them. If no words come back, nothing is sent, the box says so, and the recording stays in your voice history.

• Pick your microphone from the input picker if you have more than one.
• Choose a reply language: Auto answers in your interface language, or pin it to English, Spanish, Portuguese, French, German, Italian, Japanese or Chinese. Deska remembers your choice after a restart.
• Watch your balance: the voice minutes you have left this month are shown in Settings under Account, and on your account dashboard on the web.
• Names come out spelled right: when speech to text mishears a name Deska knows, the words land fixed, so Cloud Code becomes Claude Code and Dets, Detz or Dext becomes Dex. Devs is also the word for developers, so it becomes Dex only where the sentence names one assistant, like Hey devs or Devs, open a terminal, and stays as you said it in the devs or for devs.

Dex can read its answers back to you. In the audio popover you pick an agent voice: Cove (calm and neutral), Ember (warm and expressive) or Sage (crisp and direct). Each has a preview, and your choice sticks across sessions.

There is nothing to configure. Synthesis runs on Deska’s managed voice service and draws on the same voice-minute allowance as dictation. An answer is spoken in one voice from start to finish. If the voice service is unavailable when an answer starts, because you are signed out, out of minutes or offline, Dex reads that answer in your operating system voice, which is functional but flat. If one sentence still fails after a second try partway through an answer, Dex skips it rather than switch voices, and the sentence stays on screen.

To cut a reply short, use Stop talking on the reply card. It silences the current answer only: the answer stays on screen, the turn is not cancelled, and the next reply is spoken as usual. Muting in the audio popover is the other thing, a preference that keeps Dex silent until you turn it back on.

Dex stops talking where its suggested next steps appear. A line written right after those buttons is almost always a recap of the answer, so it stays on screen without being read out; if Dex goes back to work after them, it talks again about what it finds.

While you are talking to Dex, a soft, Apple-Intelligence-style glow breathes around the edges of the window and pulses in time with your voice: a gentle, ambient signal that it is listening, without a modal stealing the screen. It reacts to your microphone level in real time, reads best against Deska’s dark themes, and fades out the moment the exchange ends.

Everything you say to Deska is kept, dictations and requests to Dex alike. The Voice panel lists it, newest first, and every entry says where its words went: Sent to Dex, Typed into a new task, Typed into app.tsx, Pasted into the app you were in. When nothing could take the words (no cursor anywhere, or a box that closed before they came back), the entry says so, and Deska tells you at that moment too.

• Search at the top finds entries by their words or by where they went, so “dex screenshot” finds the screenshot request you sent to Dex.
• Copy puts an entry’s words on the clipboard.
• Delete shows on hover or keyboard focus; right-clicking an entry offers both.
• Clear history, at the end of the list, deletes everything once you confirm.

The panel only keeps the history. It does not record, and it does not type entries anywhere: pressing inside it puts the focus on the panel itself. To put past words back where you are typing, press the small arrow beside Dictate on Dex, whether Dex is resting or open for typing. It lists your recent entries, newest first: press one to put it back at your cursor, or Copy it. Open as a panel at the bottom opens the full Voice panel.

A dictation is never lost when its words do not come back. The moment you stop talking, the recording is saved on this computer and an entry appears in the list. If transcription fails, returns no words, or you stop it halfway, the entry stays marked Not transcribed (or Partly transcribed, keeping the words that did arrive), says why, and offers Retry transcription. Once a retry works, the words fill the entry and the saved recording is deleted. Deleting the entry deletes the recording too.

Transcription runs on Deska’s servers, so there is nothing to set up. It is measured in minutes: 10 a month on Free, 500 on a paid plan. Spoken replies draw on the same allowance.

Every plan comes with a voice allowance, and voice runs on Deska’s keys, so there is nothing to configure. A provider key stored by an earlier version is never spent on transcribing what you say.

The same voice experience travels to the companion app on the web and your phone, connected back to the machine where your project lives. It is the easiest way to check on a long task when you are away from your desk.

• Hold Dex at the bottom right of the screen and speak. Dex’s screen rises behind your thumb and he listens in the middle of it, his mouth moving with your voice. Release to send. Tap him instead to type.
• Slide up to the lock to keep him listening hands-free, then tap send. Slide left to cancel: the screen goes back down and you are where you were.
• The chip at the top names the computer the turn runs on, with its system’s logo. Tap it to pick another one for this conversation.
• While Dex works, each step shows an icon for what it touched: a terminal, a browser, an agent. His answer stays on the screen until you ask again or close it, with an Open button for each panel or task it names.

Spoken answers work as they do on the desktop: one voice per answer, read a sentence at a time, and Dex stops talking where his suggested next steps appear.