Does Keebye work inside a terminal running Claude Code?

Yes. Keebye recognizes terminal focus for its insertion plan. Paste is the default; enable type mode for SSH, tmux, non-QWERTY layouts, or other targets where clipboard insertion is unreliable.

Does Claude Code or Anthropic ever see my audio?

No. Transcription happens on your Mac and the audio never goes anywhere. What reaches Claude Code is the finished text, the same as if you had typed it.

Does it work offline?

Yes. Once the one-time model download finishes, transcription runs fully local with no connection needed — though Claude Code itself of course needs one to reach its model.

How do I start and stop dictating?

Hold Right Command while you speak and release to drop the text into the session — or tap once to toggle for longer prompts. Esc abandons a dictation mid-flight. Fn and Right Option are the built-in alternative hotkeys if Right Command clashes with something.

Can it learn project-specific terms?

Add recurring repo, branch, and framework terms to the custom dictionary to improve normalization, then review identifiers before sending.

Voice Dictation for Claude Code on macOS

Drive Claude Code by voice: hold a hotkey, speak the prompt, and Keebye types it into your terminal — on-device, offline, terminal-aware.

Why dictate prompts to Claude Code

Lane one is running tests. Lane two has finished a refactor but renamed files it was told to leave alone. The next unit of work is not a code line; it is a redirect: keep the extraction, revert those renames, and rerun the suite. That formed instruction is faster to speak than to retype while you are still reviewing lane one.

This is exactly the scenario Keebye was built for: builders running several agent workstreams at once, narrating instructions to each one instead of typing them. When three terminals each have a Claude Code session waiting for direction, speaking the next prompt keeps you moving between them.

How it works

Focus the terminal running Claude Code, hold Right ⌘, and talk. Release, and the transcript lands right in the Claude Code input, ready to send or edit.

A terminal is a special case for getting text in, and Keebye treats it as one. Paste mode uses a terminal-aware insertion plan. For SSH, tmux, non-QWERTY layouts, or a paste-hostile target, enable type mode so Keebye sends Unicode keystrokes without using the clipboard.

Everything runs on your Mac: Parakeet, the default English model, transcribes on-device, and once its one-time download is done, no network is needed and nothing is phoned home. Your audio stays local — Claude Code just sees text, the same as if you'd typed it.

Setup in two minutes

Setup is short: install the app, approve the Accessibility and microphone prompts, and choose a hotkey. Right ⌘ is the default; Fn and Right ⌥ ship as alternatives. With launch-at-login on, Keebye is up before your first session is.

Then give the custom dictionary a minute: repo names, tool names, and project jargon you repeat all day. It reduces routine corrections, but identifiers still deserve a glance.

Limits, honestly

One expectation to set: Keebye transcribes when you release the key, not while you talk. There are no live streaming captions — you speak, let go, and the full prompt appears at once.

Keebye is macOS only.

Return to lane two: focus its terminal, speak the redirect, review the text, and send it while lane one continues running. That handoff is the workflow in Driving Claude Code with your voice. If you are comparing local-first tools first, read Keebye vs Superwhisper.