Skip to content

Driving mode

Driving mode turns a session into a car screen. The phone reads short replies aloud, shows big tiles, and takes spoken commands. It doesn’t make phone use safe. Keep the phone in a mount, plugged in, and follow your local rules.

  1. Open a running session on your phone.
  2. Open the session menu and choose Driving mode.
  3. Pick a Speed of 0.9×, 1×, 1.15× or 1.3×.
  4. Optionally pick a Reading voice and turn on Captions to show the line being read in large type. Turn on Listen after replies if you want the mic to open after each reply.
  5. Choose Start. The phone says “Driving mode on.”

Reading voice and Listen after replies only show when the browser can offer them. Listen after replies needs speech recognition. If the browser can’t read aloud at all, the sheet says so and approvals still work with taps.

The phone keeps its screen on while driving mode runs, where the browser allows it. Driving mode is for one session at a time. Reloading the app during a drive brings you back to the same screen.

A bar on top shows one state word and Exit. Under it, a two-line slot shows what the phone says. The state word is one of Ready, Listening, Sending, Mic blocked, Working, Speaking, Your answer, When parked, Reconnecting or Switching. When parked means the agent asked a question and waits.

The home screen has a large Now tile with the session, its number and its computer. Beside it sit Talk, Sessions, Waiting, Repeat, Skip and Stop. Every tile is at least 64 px tall. There is no transcript and no typing. The screen follows the app’s light or dark theme.

  • Repeat reads the last line again.
  • Skip drops the line being read and moves to the next.
  • Stop silences the line and everything queued.

A headset or steering wheel button works too, on browsers that support it. Play and pause are Talk while the phone is quiet and Stop while it speaks. Next skips and previous repeats.

  • The agent’s own spoken summary. While you drive, the computer adds one paragraph to each prompt you send, asking the agent to end its reply with a spoken summary in plain sentences, as long as it needs to be. The phone reads that line. Your transcript in the app shows only what you wrote, and the agent’s own history keeps the extra paragraph. Slash commands and messages from other agents don’t get it.
  • If no summary arrives within 8 seconds of the turn ending, the phone reads the first two sentences of the reply, cut at 200 characters, with code, paths and links shortened.
  • A turn that only used tools gets a count such as “Edited two files, ran one command”, or “Done. Details are on screen.” when the phone didn’t see the work.
  • A new approval, read from the approval itself. The line comes from the call, never from agent text, for example “The agent wants to run pnpm test.”
  • A fixed line when the agent asks a question: “The agent has a question. It needs the screen.” The question stays on screen.
  • A fixed line when the session ends, the agent stops unexpectedly, a turn fails, or Panic stop runs. Error text from the provider is never read.

Reasoning, tool output, sub-agent replies and messages between agents are never read. After a reconnect the phone reads only what it hasn’t read yet.

Say status at any time to hear how long the turn has run, or that nothing is running.

Press Talk and say one thing. A whole utterance that matches one of these is a command:

Say (English) Say (Spanish) Does
stop para Silences the phone
repeat repite Reads the last line again
skip siguiente Skips the line being read
faster más rápido Next speed up
slower más lento Next speed down
status estado Says what the turn is doing
cancel cancela Drops a prompt before it goes

Anything else, up to 300 characters, is a prompt. The phone reads it back, then sends it after 3 seconds unless you say cancel, press Cancel, or press Send to go sooner. Send stays inactive until the read-back has finished. A longer prompt is refused with “Too long for driving.”

The mic keeps listening through short pauses. It reads your prompt back after 2.5 seconds of silence by default. Change that in Computer settings, Voice, under Wait after you stop talking: 1.5, 2.5 or 4 seconds. The same page holds the reading voice, the Voice language and the Reading speed.

Voice language is Auto (app language), Español (es-ES) or English (en-US). It sets the language of dictation, voice commands and spoken replies, so a Spanish speaker in an English app can get Spanish recognition. It is saved on this device. A reply is read in the language it is written in, with a matching voice, and the voice language settles messages that don’t show their own. Reading speed runs from 0.8× to 1.6× and multiplies the speed picked for the drive. Before reading, the phone drops ids, paths, links and code, and says symbols such as & and -> and dates as words.

With Listen after replies on, the mic opens for 6 seconds after each spoken reply. A prompt then goes only when you say send (in Spanish, enviar). Silence drops it. After 2 quiet windows in a row the mic waits for a Talk press. Turn off in the bar ends it. It never answers approvals or questions.

If the phone can’t send because the connection is down, it says “Not sent” and sends nothing later.

A new approval takes over the screen. It shows what the agent wants in large type, with Talk, Deny and Approve. A tap on Approve or Confirm waits 700 ms after the approval appears, so a tap meant for the last one can’t answer this one. Deny answers at once. Both are inactive while the link to the computer is down, so an answer never queues and lands minutes later. After you answer, both stay off until the computer settles it.

Which answer an approval takes depends on the call:

  • A plain tool call, such as an edit inside the workspace, takes one tap on Approve. By voice it takes one phrase.
  • A call with a risk label (Destructive command, Network access, Git push and the others), a guard, an MCP call with arguments, a path outside the workspace, or a shell command off a short safe list takes a second step. Tap Approve, then Confirm within 8 seconds, with the risks listed. By voice it takes two phrases.
  • A write to a config file that decides what later commands run, a command that changes who controls the computer, a file edit outside the workspace and a permission mode upgrade can’t be approved by voice. Voice can still deny them. You tap Approve then Confirm.

To answer by voice, press Talk. The phone reads the approval, then names a phrase, for example “Say: approve blue.” The colour is random for each approval: blue, green, orange, purple, silver or yellow. For a two-step approval it says the risks and then a second phrase such as “confirm seven”, with a number from three, five, six, seven, eight or nine. You have 60 seconds between the two phrases. In Spanish the phrases are “aprobar azul” and “confirmar siete”.

Only the whole phrase counts. A wrong colour or extra words match nothing. Two misses, 8 seconds of silence or four mic openings in a row leave the approval for the screen. Say no, deny or stop to deny, and later to leave it. At the second step, a wrong phrase cancels with no retry.

The computer checks the same rules and refuses a spoken answer that carries fewer phrases than the approval needs. A deny sends the agent the note “Denied by voice while driving.”, or “Denied while driving.” for a tap.

A phone that can’t approve, such as one paired as a Viewer, shows “This device can’t answer approvals right now.” The Talk tile on an approval only shows when your browser has speech recognition and your computer takes spoken answers.

Say menu, or tap Sessions, for a numbered grid of up to 9 sessions across your computers. The current one comes first, then those that need you. Say a number such as “number 3”, or “switch to” and a session name. Say go back to return.

Waiting shows how many sessions need you. When another session needs you, a banner names it and the kind of ask, with Yes, No and a 10 second ring. It never switches by itself. If you don’t answer, it asks once more, and then a badge waits on Sessions. After a switch from a banner, Waiting becomes Go back. A spoken yes or no on that banner never answers an approval.

  • The phone must stay on with the app in front. A locked screen or a background app pauses driving mode, because phones suspend web apps. When you return, the phone renews its hold on the session and reconnects.
  • The computer stops asking for spoken summaries 15 minutes after the phone last renewed its hold.
  • Driving mode speaks and listens in English or Spanish, following the app language unless you set a Voice language.
  • If your browser has no speech recognition, Talk and Listen after replies don’t show. You still get spoken replies, the tiles and tap approvals.
  • Some browsers send microphone audio to their own speech service.

Choose Exit in the bar. Pressing Back or opening another session also ends it. If the screen fails to load, a Close button ends driving mode and returns the session. Leaving the screen tells the computer to stop adding the paragraph to your prompts.