How to use MindClerk.

MindClerk is in a private beta; downloads open together with accounts and plans. Once it is installed, a short setup asks for the Microphone and Accessibility, downloads a speech model and offers Gemma, a free model that cleans up your dictation on your Mac. What it needs, and why.

MindClerk lives in the menu bar and in a thin bar at the bottom of the screen. Closing the window only hides it; the shortcut stays live. The main window has your history (Dictation), meeting notes (Notetaker), Insights, the Dictionary and Transforms, with Settings and Help at the bottom of the sidebar.

1. Click into any text field — a note, an email, a chat, the terminal.
2. Hold ⌥ Space. The bar turns into a moving waveform and a soft sound plays. Speak; pause as long as you like.
3. Let go. The bar shows Transcribing, then Polishing, and the text appears at the cursor.

For long dictations use hands free: tap ⌥ ⇧ Space to start and again to stop; it also stops by itself after 2.5 seconds of silence. Esc cancels. A recording stops by itself after three minutes, and a tap shorter than 0.3 seconds is ignored.

The text is fitted into what is already there: no capital in the middle of a sentence, no double spaces, and a dictated list pastes as a real list. Every dictation is kept in your history and on the clipboard, so a paste that went to the wrong window is never lost.

The bar sits at the bottom of the screen at all times and never takes focus from the app you are typing in.

Point at the bar and three buttons rise out of it: the globe picks the language of every dictation from now on (or Detect automatically), the microphone starts the Notetaker, and the wand picks the transform that runs after each dictation. Drag the bar anywhere — it remembers the spot — and click it to open the main window. To hide it for a while, use the menu bar item.

The defaults. Change any of them in Settings → General (they are also listed in Help → Essentials). A shortcut is taken over system-wide, except Esc.

• ⌥SpaceHold to talkRecords while held; let go to paste.
• ⌥⇧SpaceHands freeTap to start, tap to stop; stops on silence too.
• EscCancelStops whatever is in progress.
• ⌥⇧ECommand modeSay what to do with the selected text.
• ⌥⇧ZUndo AI editYour own words back; press again to redo.
• ⌥⇧PPolish againCleans up the last dictation once more.
• ⌃⌘VPaste last dictationPastes the most recent text again.
• ⌥SScratchpadOpens or closes the floating editor.
• ⌥⇧NMeeting notesStarts or stops notes of the call.
• ⌥⇧AMeeting assistantAsk about your meetings; the answer is typed.
• ⌥1Clean up selectionRuns the transform on selected text.
• ⌥2Technical / terminalSpoken syntax becomes real syntax.
• ⌥3Professional emailA polished email from the selection.

AI cleanup runs on every dictation: it adds punctuation and capitals, drops fillers and false starts, applies your self-corrections (“Tuesday — actually Thursday” → “Thursday”) and turns a dictated list into a list. It does not rewrite what you said, and if the model ever strays from your words, what you said is pasted instead.

Cleanup is one transform — a named instruction for text. Nine come with the app: Clean up, Technical / terminal, Professional email, Code, Chat, Bullet points, Translate to English, Prompt engineer and Transform selection. Run one after every dictation (the bar’s wand, or Settings → Transforms → Apply after dictation), or on selected text with its shortcut — ⌥ 1 Clean up, ⌥ 2 Technical, ⌥ 3 Professional email.

Make your own on the Transforms page: a prompt for a model, or a list of exact rules that needs no model at all. Give it a shortcut and, if you like, a model of its own.

Didn’t like an edit? ⌥ ⇧ Z puts your own words back; press again to redo. ⌥ ⇧ P polishes the last dictation again, and ⌃ ⌘ V pastes it once more.

Select text in any app, hold ⌥ ⇧ E and say what to do: “make this shorter”, “translate to English”, “turn this into a bulleted list”. Let go and the selection is replaced. With nothing selected it works on your last dictation and puts the result on the clipboard.

Modes change what happens in particular apps. The first that matches the app in front wins:

Edit them in Settings → Models & AI → Modes: match by app or by window title, turn cleanup off for one app, or choose how the text is inserted. Some apps should never see MindClerk — games, password managers, remote desktops: add them under Settings → System → Apps where MindClerk stays off.

The dictionary teaches the speech model your names and terms, and fixes the ways it mishears them. Snippets turn a spoken phrase into longer text — say “my address” and it is written out — with placeholders for the date, the time, the app and the clipboard, and an optional shortcut to paste one without dictating.

The scratchpad (⌥ S) is a floating editor for when there is no text field to dictate into, or you want to think out loud first. Its buttons run a transform on the selection; the text is kept between launches.

The Notetaker records your microphone and the sound your Mac plays — the other side of the call — writes a live transcript, and when the meeting ends adds a title, a summary and action items, in the language the meeting was held in. Nothing joins the call: no bot, no guest.

Start notes in any of these ways:

• ⌥ ⇧ N, from any app.
• Meeting detected — a pill above the bar when a call starts in Zoom, Meet, Teams, Webex, FaceTime, Slack, Discord, WhatsApp, Telegram and more, in the app or the browser.
• A reminder before a meeting on your calendar, with Join + take notes.
• New Note on the Notetaker page, or the menu bar item.

Speakers are told apart on your Mac, and names are taken from what Meet, Zoom or Teams shows as speaking. Action items collect on the Meeting Tasks page. Ask your notes on the Notetaker page, or hold ⌥ ⇧ A in any app to ask about your meetings and have the answer typed at the cursor. Your notes can also be read by Claude or Cursor through a local MCP server.

Recording the other side needs macOS 14.6 and the Screen & System Audio Recording permission. In many places you should tell people you are taking notes — the Notetaker keeps a consent message ready to paste into the call’s chat.

Every dictation is saved on your Mac, as you said it and as it was cleaned up, with its recording. The Dictation page lists them by day: search, copy, replay the audio, or run one through a different transform. Turn recordings off, or history altogether, in Settings.

Insights turns that history into a few numbers: words dictated, your speed against typing, your streak, words per day, where you dictate and how long the text takes to appear. Nothing on these pages leaves your Mac.

Speech is recognised on your Mac by Parakeet (fast, 25 languages) or Whisper (99 languages), and cleaned up by Gemma, also on your Mac. Choose them in Settings → Models & AI.

Prefer a cloud model? Pick OpenAI, Anthropic, Gemini, Groq or OpenRouter for text, or OpenAI, Groq, Deepgram or ElevenLabs for speech, and paste your own API key — it goes to the macOS Keychain and the provider bills you directly. Ollama works too, on your Mac. Without a key, the app uses the model on your Mac.

• What leaves your Mac (Settings → Data & privacy) says, for your current settings, what each stage sends and to whom. Out of the box every line reads “Stays on this Mac”.
• Privacy mode keeps nothing from a dictation on disk: no history, no recording.
• Hide from screen sharing keeps the window, the bar and the scratchpad out of screen shares and screenshots.
• Export, Import and Reset your setup; delete history and notes from the app.

There are no analytics, no telemetry and no crash reporter. See the privacy policy.

• The shortcut does nothing, often after an update. System Settings → Privacy & Security → Accessibility: switch MindClerk off and on again, or remove it and add it back.
• The text did not appear. It is on the clipboard and in your history — paste it with ⌘ V. Some apps refuse pasted text; a per-app mode can insert it another way.
• A short phrase came out in the wrong language. Pick the language with the bar’s globe, or say a full sentence — detection is surer with more to go on.
• No “Meeting detected” pill. Check Settings → Notetaker → Offer to take notes when a call starts. Browser calls need Accessibility.
• Still stuck? Help → Something is not working in the app, or write to [email protected].