Dictation
Say a prompt instead of typing it. Every agent prompt in Agentastic has a microphone: talk, stop, and the words land at the cursor, transcribed on your Mac.
Every prompt you write to an agent has a microphone beside its send button. Click it, say what the agent should do, and click again: what you said lands at the cursor, ready to edit before you send it. It works the way it does in the companion app, with the same speech engine.
Dictate a prompt#
- Click the microphone, or press Option+Cmd+V with the cursor in the prompt. The button turns into a red stop button.
- Talk. The button widens into a capsule with a level meter, which is how you know your Mac hears you. Nothing is written while you talk.
- Click the stop button, or press Option+Cmd+V again. A spinner shows briefly while the last words settle, and then the text appears.
What you said goes where the cursor is, replacing the selection if there is one, with the spaces it needs on either side. It's a single edit: Cmd+Z takes it back, and you can edit the words like anything you typed. To throw a recording away instead, press Esc in the prompt.
You don't have to stop first to send. Press Return or the send button while you're still talking, and the message goes out as soon as your last words are in it, including a prompt you only spoke. If nothing was heard, nothing is sent and the prompt says Didn't catch that. Press Return again to send what's already there.
Only one prompt listens at a time. Starting dictation in a second prompt finishes the first one and keeps its words. Closing a prompt stops its dictation. In Agent Home and Quick Task, what you said still lands in the task's draft.
Where you can dictate#
| Prompt | Where the microphone is |
|---|---|
| Agent Home and Quick Task | Between + and the send button |
| Chat follow-ups, in every pane and the sidebar | Between + and Stop or Send |
| Quick Edit (Cmd+Shift+K on a selection), and its follow-up field | Before Edit |
| Comments on a Markdown selection | Before Send |
| Browser design feedback | Before the send button |
| Manager Agents: messages, comments, and answers | Before the send button |
| Answers to an agent's question in Chat | Before Send |
| The prompt of a Scheduled Agent | Beside Edit / Preview |
In the first four, the shortcut and Esc work too. In the others, click the microphone, and the words are added at the end of what's already there.
Privacy and languages#
On macOS 26 and later, your Mac transcribes with the on-device model behind Notes and Voice Memos transcription. The audio never leaves your Mac. The first time you dictate in a language, macOS may download its speech model once; a ring around the stop button shows the progress.
On macOS 14 and 15, dictation uses Apple's older speech recognizer. It runs on your Mac where it can, and otherwise Apple's servers transcribe. That's why macOS asks for Speech Recognition as well as the Microphone the first time. On macOS 26, only the microphone is asked for.
Dictation listens for the first of your Mac's preferred languages. To speak another one, choose it under Settings → Agents → Composer → Dictation language.
Names from the prompt's context (its agents, file, and repository) and common coding words such as "pull request" go to the recognizer as hints. The keyboard's dictation model and the older recognizer use them. macOS 26's newest model mostly transcribes what it hears, so check names before you send.
Turn it off#
Turn off Settings → Agents → Composer → Dictation to hide the microphone from every prompt and free the shortcut. To give Dictate into Prompt another key, change it in Window → Keyboard Shortcuts. If Option+Cmd+V was already bound to something else when you updated, that binding is kept and Dictate into Prompt starts out disabled.
When it doesn't work#
- Microphone access is off for Agentastic.dev. Click Open System Settings, turn on Agentastic.dev under Privacy & Security → Microphone, and try again. Speech Recognition works the same way.
- No microphone is available. Connect one, or choose an input in System Settings → Sound.
- Didn't catch that. The recording was silent. Check that the level meter moves when you talk; if it doesn't, the wrong input may be selected in System Settings → Sound.
Agents running in a terminal keep their own voice features. Claude Code's voice dictation, for example, uses the same microphone permission, and an agent you switch to Chat (Cmd+I) gets the composer's microphone.