Voice simulation
Say it, and the router picks the agent
- KeyboardYou are in the Inbox and the answer is 1, 2, or 3. That is faster than speaking.
- VoiceThe instruction is a sentence, or you want to point at a button on screen instead of describing it.
The shortcut and the router
The ⌥ double-tap works even when doki is not in front. That is why it asks for Accessibility: not to read the whole screen on its own, but to hear the shortcut and, when needed, type into the agent’s window.
The router does not open a generic chat with a doki model. It chooses among the agents already running. If none of them is on that task, it can start one. The criterion is the task, not “whichever I used last”.
Point instead of describing
You drag a rectangle on the screen and say what should happen. The agent receives the image and the sentence. It fits “this button is cut off on mobile” or “the empty state of this list says nothing”, cases where text alone forces the agent to guess.
Screen recording is the permission that makes it possible. Without it, Talk still works for dictation. Without the microphone, Talk still works for typing.
Voice notes inside a thread
Besides the router, each conversation accepts a voice note. doki transcribes it and sends it as text, so the agent does not depend on an audio attachment. On the iPhone it is the same gesture: you dictate and the note lands in the thread that was already waiting.
That is for when you already know which session it is. The router is for when you do not want to find it. Both paths end in the same Inbox, described in the Inbox guide.
Voice with several agents open
The obvious risk is sending the message to the wrong agent. The router uses the context of who is already on that task. If it is ambiguous, look at the destination before the two seconds pass, and cancel with Esc if it is the wrong one.
Speaking does not replace worktrees. Voice does not isolate files. If Claude and Codex edit the same tree, voice only speeds up the conflict. Isolation is in several agents and in worktrees.
What actually gets recorded
Dictation uses the Mac’s speech recognition. The capture is the region you mark, not a continuous video of the desktop. Those pieces travel to the agent as context for that reply, the same as if you had pasted them into the terminal.
They do not become a surveillance archive. If the microphone or screen permission is missing, the matching feature does not start. They are granted on first use and can be removed in System Settings.
Frequently asked questions
Does doki listen all the time?
No. Talk opens with the shortcut. Outside of that, the microphone is not capturing for doki.
Does it work in Spanish?
The app is in Spanish and English, and it follows the system language. Dictation depends on the speech recognition you have installed on the Mac for that language.
Can I dictate from the iPhone?
Yes. The thread accepts voice notes, and you can also type or pick an option. The ⌥ double-tap router is on the Mac, because the shortcut lives on the keyboard.
What if dictation gets it wrong?
You have about two seconds to press Esc. The message does not leave, and you can dictate again or type it.
