Dictation Not Working With Your Coding Agent

Before you change any settings, work out which of two different things is broken, because they fail for different reasons. Operating system dictation types text into whichever field has focus and works across every app. An in-app voice mode records audio and sends it to a service over the network. Say a sentence into a plain text editor using OS dictation: if the words appear, your microphone and its permissions are fine, and the fault is in the agent or the field you are typing into.

Have I granted microphone permission to the right thing?

This is the most common cause and the most misleading one, because the thing that needs permission is often not the thing you are looking at. For a terminal-based agent such as Claude Code or Codex, the microphone permission belongs to the terminal application, not to the agent running inside it. Granting access to one terminal does not grant it to another, so moving to a different terminal app can break dictation that worked perfectly yesterday.

On macOS this lives in the privacy section of System Settings under Microphone, and the application has to be listed there and switched on. The prompt appears once. If it was dismissed, it does not come back, and nothing in the app will tell you why it has gone quiet. In a browser the equivalent is the site permission for that tab: block the microphone once for an origin and the browser keeps refusing without asking again.

Is the right microphone actually selected?

Check that the input level meter moves when you speak, in the operating system sound settings rather than inside the app. If it does not move, the selected input is not the microphone you are speaking into. Bluetooth headsets are the usual offender: they present differently depending on the mode they are in, and a switch between modes can change the audio quality or drop the input entirely. A wired microphone removes a whole class of problem here.

Also rule out the dull ones: a physical mute switch on the headset or its cable, a mute at the operating system level, and another application, usually a video call, holding the microphone.

If dictation works but keeps losing your first word, that is timing rather than a fault. Push to talk starts recording at the moment of the press and the audio path takes a beat to open, especially over Bluetooth. Press, pause, then speak.

Dictation works elsewhere but not in the agent's input box. What now?

Then the field is the problem. OS dictation inserts text into standard text inputs. Some interfaces are not standard text inputs: certain terminal user interfaces, canvas-based editors and custom prompt boxes accept keystrokes but not inserted text, so your speech goes nowhere visible. The test is to dictate into a plain text editor and then paste the result into the agent. If pasting works and dictating does not, the input field is the thing rejecting you, and no amount of microphone fiddling will help.

What if the app's own voice mode is the part that fails?

In-app voice modes depend on a network round trip and a live session, so they fail in ways local dictation does not. If the button changes state, the level meter moves and no transcript ever arrives, that points at the service rather than your machine. Reloading the session is worth one attempt. After that, checking the tool's own status page beats another hour of local troubleshooting.

Would dedicated hardware fix any of this?

Only part of it, honestly. A hardware push-to-talk key sends a keystroke, so it can only trigger dictation that already works. It will not grant a permission or pick an input device for you. What it does fix is the ergonomic half: holding a shortcut with one hand while you read a diff with the other, or hunting for a key combination you keep missing. Our Whisper Key at $87.90, Voice Pad at $79 and Vibe Pad at $59 are built around that press, and the two supplied with a microphone at least give you one known, wired input device instead of whatever Bluetooth decided to do this morning.

If the real problem is dictating quietly rather than dictating at all, dictating code prompts in an open-plan office is the better starting point.