Which AI tools does this work with?

Keebye works with ordinary editable fields in browsers, desktop apps, editors, and terminals; it does not need a per-tool plugin and deliberately refuses secure fields.

Does the AI provider get my audio?

No. The speech-to-text step happens on your Mac before a single character reaches the tool. The model on the other end receives plain text and cannot tell it came from your voice.

Is spoken text good enough to send as a prompt?

Usually, yes. Rule-based cleanup fixes obvious spoken-language artifacts. The optional on-device polish uses a heuristic that catches severe token loss, expansion, and repetition, but it cannot guarantee meaning or negation. Review critical qualifiers before sending.

How long a prompt can I dictate at once?

Tap-to-toggle supports multi-paragraph briefs without holding the key. Practical duration depends on your Mac and available memory, so review long prompts before sending.

What about code, file names, or library names in my prompt?

Add recurring repo names, package names, and API terms to the custom dictionary. It improves how those terms are normalized, but you should still review identifiers before sending.

Voice Dictation for AI Prompts on macOS

Speak complete AI prompts instead of typing them: hold a hotkey, say the request, and Keebye inserts editable text into the focused app.

Why speak your prompts

A good prompt is a paragraph, not a search query. Imagine asking an agent to move retry logic into its own module, preserve the public API, add one failure test, and leave generated files alone. The thought is already formed; turning all four constraints into text is the slow part.

Say it instead. Speaking a multi-sentence request is closer to how the idea actually arrives in your head, and it comes out in one breath rather than one keystroke at a time. This is the heart of what Keebye is for: your voice becomes the input to AI tools that accept text, whether that is Claude Code in the terminal or a chat window in the browser.

How it works

Click into the prompt box, hold Right ⌘, and describe what you want the model to do. Let go and the full request appears at the cursor, ready to send or refine. For a longer brief, tap the key once to keep dictation open, talk through the whole thing, and tap again to stop.

The transcription is on-device — Parakeet, the English default, downloads once and then runs without a connection. Audio is not uploaded. The finished text can be saved in Keebye's local, text-only history for 30 days by default; you can turn history off if you want nothing retained.

The identical hotkey works in ChatGPT in your browser and in a native app the same way, so switching tools costs you no new habit.

Setup in two minutes

Allow Accessibility and microphone access on first launch, then pick your trigger key — Right ⌘, Fn, or Right ⌥. Enable launch-at-login and Keebye is waiting the next time a prompt box is in front of you.

Spend the last minute on the custom dictionary. Prompts are dense with proper nouns — model names, product names, the libraries you reference constantly — so teaching those once reduces repetitive corrections. Review file names and identifiers before sending.

Limits, honestly

Keebye writes the prompt; it does not run it. There is no integration that reads a tool's state or clicks send for you — it drops text where your cursor sits and you stay in control of the conversation, the same as you would dictating into Cursor's composer.

The text arrives when you release or toggle off the hotkey, not word by word as you talk — transcription is batch, not streaming. And Keebye is macOS only.

Return to the retry brief: speak the context, four constraints, and expected test in one pass; read the resulting paragraph; then decide when to send it. For the fuller workflow, read Driving Claude Code with your voice. If you are choosing a dictation architecture first, start with Keebye vs Wispr Flow.