Skip to content
MelMel
Agent

Voice input

Dictate prompts instead of typing.

Sometimes it's faster to say the prompt than to type it. Click the microphone, speak, click again, and the words land in the box you were typing in.

Where the mic lives

  • The agent input — a mic button next to send. Hide it with Voice input in Settings → AI if you never use it.
  • The terminal's toolbar — the same mic icon, so you can dictate a command.
  • The tool belt — a Voice button whenever an embedded CLI agent is running, so you can dictate into Claude Code or Codex too.

Click to start: a Listening… overlay appears with a live level meter — it moves with your voice and stays flat if the microphone is muted or silent, so you can tell at once whether Mel hears you. Click again to stop; the text is pasted into the focused input (into a CLI agent through a bracketed paste). On a Mac, dictation works since v0.1.56.

Where your audio goes

The first time you dictate, Mel asks: "Send this recording to Mel's speech-to-text?", and explains what happens before anything is sent. Choose Allow cloud transcription to go on.

  • Mel records only while the mic button is active. When you stop, that recording is sent to Mel's servers and on to Groq, which turns it into text and sends the text back.
  • Groq's published terms say it does not retain customer data for inference requests by default; it may keep a log for up to 30 days while troubleshooting an error or investigating abuse, and it may not train on your inputs or outputs without permission, which Mel has not given.
  • Mel stores no audio. The recording is held in memory, uploaded once, and dropped; the text goes into your input box and nowhere else.
  • Voice is metered in audio seconds against your plan's daily voice allowance and your credits — see Settings → Billing and Usage.

Keep it on your computer

Prefer nothing leaving your machine? Set MEL_VOICE_CMD to a local transcription command, such as a whisper.cpp wrapper. Mel checks it first, gives it the recording, and pastes what it prints. It needs no account, no consent dialog, and nothing is uploaded.

# Example: a local transcription script of your choosing
export MEL_VOICE_CMD="$HOME/bin/transcribe.sh"

Limitations

  • No keyboard shortcut: the mic is clicked.
  • No dictation inside ssh, vim, less, top and other full-screen programs (those panes have no tool belt).
  • No live, word-by-word transcription: the text arrives when you stop.
  • No language picker.

FAQ

Does Mel listen all the time? No. It records only between your two clicks.

Can I use dictation without a Mel account? Yes, with MEL_VOICE_CMD. Cloud transcription needs a signed-in account, because it is metered.