Skip to content
Agent

Voice input

Dictate prompts instead of typing.


Sometimes it's faster to say the prompt than to type it. Mel has a dictation flow wired end-to-end — with one honest caveat up front: built-in transcription isn't shipped yet. Today, voice input works through a pluggable local hook you provide; a bundled local speech model is on the roadmap.

Where the mic lives

  • The agent input footer — a mic button next to send. It can be hidden via the voice toggle in Settings if you never use it.
  • The tool belt — a Voice button whenever an embedded CLI agent is running, so you can dictate into claude or codex too.

Click to start: a "Listening…" overlay appears with a pulsing mic. Click again to stop; the transcription is pasted into the focused input — the agent's, or the CLI agent's via bracketed paste.

Local-only by design

Mel never streams your microphone audio to a cloud speech-to-text service. That's a deliberate privacy stance, not a missing feature: transcription is meant to happen on your machine.

Today that works through an environment hook. If MEL_VOICE_CMD is set, stopping dictation runs your command through the shell and uses its stdout as the transcribed text:

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

Your script can do whatever you like — record and run a local speech model (whisper.cpp is the natural fit), or call any pipeline you trust. Whatever it prints is what gets pasted.

Without a hook

If MEL_VOICE_CMD isn't set, stopping dictation pastes a clearly labelled placeholder telling you to configure it — never fabricated text. The UI flow (mic button, overlay, paste-on-stop) is fully functional either way, so once you wire a backend it just works.

What's coming

The planned built-in path is local audio capture feeding a bundled local speech model — same flow, zero setup, still no cloud. Until then, treat voice input as bring-your-own-transcriber:

PieceStatus today
Mic buttons, Listening… overlay, paste-on-stopShipped
MEL_VOICE_CMD local hookShipped
Bundled local transcriptionComing — not yet shipped
Cloud speech-to-textNever — by design

If dictation matters to your workflow now, the hook is a ten-minute setup with any local speech tool; if not, the mic will quietly get smarter in a future release.