Power Tools

The PowerToys of macOS. One hotkey, a whole toolbox — and the core tools never leave your Mac.

Website: powertools.geeksare.cool · Download the latest DMG

Power Tools puts a suite of power-user utilities behind a single hotkey. Hold Option + Shift, then speak or tap a letter/arrow: dictate into any app, open a Claude chat, snap and tile windows, OCR the screen, grab a screenshot, cut & paste files, transform your clipboard, or spotlight your cursor. If you've missed Windows PowerToys on the Mac, this is that — FancyZones-style snapping, Advanced Paste, Text Extractor, Color Picker, and Find My Mouse — plus on-device dictation and AI chat.

In the box: 🎙 Dictation · 💬 AI chat · ⌨️ Chord mode (multi-letter shortcuts, every menu item) · 🔀 ⌘Tab like Windows Alt-Tab · 🪟 Window snapping + Snap Assist · 🖐 Grab & Move (hold ⌃⌘, drag anywhere on a window) · 🎛 Snap palette (Moom-style) · 🗔 Saved layouts · ▦ Grid placement · 📋 Clipboard history (Win+V-style, text + images) · ⚡ Quick Capture → your own endpoints · 🔤 OCR (Text Extractor) · 🔊 Read Aloud · 📸 Screenshots · 🎥 Screen recording · 🔎 Google Lens · 🎨 Color Picker · 🗂 File cut/paste · 📝 Advanced Paste · 🖱 Find My Mouse · 🔇 Call-safe dictation

The app bundle and data folder are still named GRC Whisper / bundle ID com.grc.whisper so macOS permissions and saved keys survive the rename.

The core tools run entirely on this Mac. Dictation is on-device by default — speech recognition is Apple's on-device SpeechAnalyzer (macOS 26), text cleanup is Apple's on-device Foundation Model, and history lives in a local SQLite file. The only network calls are the opt-in AI features — AI chat, cloud cleanup, and Advanced Paste's smart transforms — and only when you invoke them, with your own API key. Little-Snitch-clean by design.

Screenshots

How it works

hold ⌥⇧ ──► mic (always-warm, 1s pre-roll) ──► SpeechAnalyzer (streams while you speak)
release ──► finalize (~0.2-0.6s) ──► Tier-0 cleanup (dictionary + fillers)
       ──► on-device LLM polish (deadline-guarded, falls back to Tier-0)
       ──► clipboard-swap paste into the focused app (your clipboard is restored)

• Hold Option + Shift (either side — it's ambidextrous) to talk, release to insert. Esc cancels. This one hotkey is also a leader: hold it and tap a letter or arrow for a different tool (see Leader chords below) instead of speaking. Configurable in Settings (Fn, Right Option, Control+Option, etc.).
• Live partial transcript shows in a bottom-center overlay that never steals focus.
• The polish pass removes fillers (um/uh), applies self-corrections ("meet Tuesday — scratch that, Wednesday" → Wednesday only), fixes punctuation, and respects your personal dictionary. If the LLM is slow or unavailable, you get the deterministic Tier-0 cleanup instead — never nothing.
• Raw + polished text for every dictation is kept in local history (menu bar ▸ Recent).

Build & install

Requires macOS 26+ (Apple Silicon) and Command Line Tools. No Xcode, no dependencies.

scripts/bundle.sh --install    # builds, signs, copies to /Applications
open "/Applications/Power Tools.app"

Package.swift exists for SwiftPM-capable toolchains, but scripts/bundle.sh compiles with swiftc directly — the CLT-only SwiftPM on this machine has a broken PackageDescription dylib.

First-run setup (one time)

TCC permissions attach to the app bundle's ID + signature, which scripts/bundle.sh keeps stable across rebuilds. Grant these when prompted (or pre-emptively in System Settings ▸ Privacy & Security):

1. Microphone — prompted on first launch.
2. Accessibility — required for the global hotkey tap and the paste keystroke. The app prompts; toggle GRC Whisper on.
3. Input Monitoring — macOS sometimes also requires this for the keyboard listener; grant it if it appears.
4. Only if you switch the hotkey to Fn/Globe: System Settings ▸ Keyboard ▸ "Press 🌐 key" → "Do Nothing", otherwise macOS opens emoji/dictation on the Fn key and fights the hotkey. The default Option + Shift hotkey needs no such remap and works on external keyboards too.

Then verify with the menu bar ▸ Permission Doctor…, or:

"/Applications/Power Tools.app/Contents/MacOS/Power Tools" doctor

After a macOS update or if the hotkey dies, re-run the doctor — event taps occasionally need Accessibility re-granted.

After every rebuild/reinstall: toggle GRC Whisper off and on in the Accessibility list. Ad-hoc signatures pin the build hash, so a rebuilt binary silently loses the grant even though the checkbox still shows enabled.

CLI

The same binary doubles as a test/administration CLI:

grc-whisper transcribe file.wav          # engine test (no permissions needed)
grc-whisper polish "um meet Tuesday scratch that Wednesday"
grc-whisper doctor                       # permission / model status
grc-whisper dict add KYAW "K Y A W,kayak"   # term + comma-separated misheard variants
grc-whisper dict list
grc-whisper history 20
grc-whisper menu-chords com.apple.finder   # chord codes for a running app's menus
grc-whisper chord-codes-test               # chord code generation self-test

The personal dictionary does two jobs: deterministic replacement of misheard variants (Tier-0, works even with polish off) and a spelling authority passed to the LLM for phonetically-near matches.

Configuration

~/Library/Application Support/GRC Whisper/config.json (also settable from the menu bar):

key	default	notes
hotkey	optionShift	optionShift (default, ambidextrous, any keyboard), fn, rightOption, rightCommand, ctrlOption, shiftCommand. fn isn't delivered by non-Apple keyboards — use optionShift or ctrlOption there.
polish	apple	apple (on-device AI), claude/openai (cloud, opt-in), basic (dictionary+fillers only), off (raw)
claudeModel	claude-haiku-4-5	cloud model when polish: claude (change to claude-opus-4-8 for max quality)
openaiModel	gpt-4o-mini	cloud model when polish: openai
localeIdentifier	en_US	any of the ~30 SpeechTranscriber locales
llmDeadlineMs	2500	polish deadline; on expiry Tier-0 text is inserted
clipboardRestoreDelayMs	600	how long the transcript stays on the clipboard before restore
maxUtteranceSeconds	120	hard cap per hold

History, dictionary, and logs live in the same folder. Delete the folder to reset everything.

Design notes (the sharp edges this app rounds off)

• First-syllable loss: the mic engine runs continuously with a 1s pre-roll ring buffer, so audio from before key-down is included. On-demand mic startup costs ~500ms and eats the first word (a chronic complaint in this app class).
• Clipboard safety: paste is a full-fidelity clipboard swap — all pasteboard item types are snapshotted and restored (images/files survive), the transcript item is marked org.nspasteboard.ConcealedType so clipboard managers skip it, and restore only happens if the pasteboard still holds our session UUID (never clobbers something you copied mid-cycle).
• Event-tap hardening: exact-match-only suppression, auto re-enable on tapDisabledByTimeout, stale-hold expiry, and a 5s health check that revives silently-dead taps (a real macOS 26 failure mode). A stuck key can never eat your spacebar.
• Secure fields: dictation into password fields is refused with a visible message (secure input also blocks the tap — nothing would work anyway).
• LLM guardrails: the polish output is sanity-checked (length ratio, refusal prefixes) and deadline-raced against Tier-0; the raw transcript is always preserved in history. Dictated text is treated as data — "ignore your instructions…" gets typed, not obeyed.

Leader chords (hold your hotkey, tap a letter)

Your hotkey doubles as a leader — hold Option + Shift and the overlay shows the menu (A ai · T text · R read · S shot · G lens · K color / C X V files · Y shelf · W palette · ← → ↑ ↓ win). Keep holding and tap a key, or just speak. Until you speak, the overlay shows the hint strip; the moment you talk it swaps to the live waveform.

Text & AI

• hold + speak + release → dictation (as normal).
• On a Zoom/Teams call? The speakers auto-mute for exactly as long as you hold the key, then restore — so the meeting audio never bleeds through the mic into your transcript. Toggle in Settings ▸ Dictation.
• hold + A, then speak → AI chat: what you say opens a streaming Claude chat window. Set the mode in Settings — native (in-app window), browser (claude.ai tab), or both (a native window with a “claude.ai ↗” button).
• hold + T → text (OCR): drag a screen region, the recognized text is copied to your clipboard. Tables come back tab-separated so they paste into Excel/Sheets/Numbers as a grid. Fully local via Apple Vision.
• hold + R → read aloud: drag a screen region, the recognized text is spoken through the system voice — hold + R again stops it. Same on-device OCR as T; nothing leaves your Mac. Better voices: download an Enhanced/Premium voice in System Settings ▸ Accessibility ▸ Spoken Content and it's used automatically.
  • Pronunciation fixes: teach it names and jargon in Settings ▸ Dictionary ▸ Read Aloud pronunciations — word + how to say it (KYAW → K Y A W, andrea → ahn-DRAY-ah). Case-insensitive, whole words only, longest entry wins; only the spoken audio changes (hold + T still copies the real text). UI edits apply immediately; the list is stored as pronunciations in config.json, where hand-edits load on the next launch.
• hold + S → screenshot: drag a region, the image is copied to your clipboard.
• hold + F → screen recording: drag the area to record (⏎ = the whole display, Esc cancels) and it starts — the cursor and your clicks are in the movie. A small ● 0:12 badge counts along; click it, or hold + F again, to stop. The .mov lands on your Desktop (or in ~/Movies/Power Tools if the Desktop isn't writable), its file goes on the clipboard (⌘V attaches it in Slack, Mail or Teams) and the Finder reveals it. You can keep dictating while it records. Settings ▸ General: turn it off, or record the microphone too for a narrated demo. It's the system recorder, driven for you — nothing leaves your Mac. If the app ever dies mid-recording, the next launch finalizes and saves what was captured.
• hold + E → whiteboard: annotate the image on your clipboard — take a shot with hold + S first, or E alone prompts a fresh region grab. Pen, arrow, line, box, circle, polygon, text, 5 colors, ⌘Z undo; , Esc cancels. (If you've assigned E to a Quick Capture connection, that wins — reach the whiteboard from the Power Ring instead.) Toggle in Settings ▸ General.

Files (with the Finder frontmost)

• hold + C copy the selected files · hold + X cut · hold + V paste — cut files are moved into the front Finder window. A keyboard cut/paste for files, like Windows Explorer.
• ⏎ opens the selection (Windows-style) instead of starting a rename — works for files and folders; Return still types normally in rename and search fields. Rename via right-click ▸ Rename or a slow double-click. Toggle in Settings ▸ General.
• Windows-style keys (Settings ▸ General): Home / End jump to line start/end in text fields (⌃ for document top/bottom, ⇧ to select); in Finder, ⌫ Backspace goes up to the enclosing folder and ⌦ Delete moves the selection to Trash; ⌃⇧⎋ opens Activity Monitor (the Task Manager reflex). Each key still does its normal job in rename/search fields.
• New Document in Finder (Settings ▸ General ▸ Set up ‘New Document’): seeds macOS's native right-click "New Document" menu with blank Word, Excel, Text, RTF, and Markdown files — the Windows "New ▸" gap. One-time setup; afterward it's Finder's own menu, no Power Tools involvement.

Windows

• hold + ← / → / ↑ / ↓ → snap the focused window to that side/edge. Tap the same arrow again to cycle sizes (½ · ⅓ · ⅔ by default; ¼/¾ or ⅕/⅘ selectable in Settings — each set includes the larger complement from the other side).
• Chain two arrows for a corner — ← then ↑ lands the top-left quarter (Moom-style); the sizes you've cycled on each axis carry into the corner.
• hold + W → snap palette: a compact panel of targets drawn as mini window diagrams — a halves row (Fill · left/center/right ½ · top/bottom ½; hold ⌥ for corner quarters + center), a thirds row (Fill · ⅓s · ⅔s), and a quarter-columns row (Fill · left · mid-left · center · mid-right · right ¼-width strips), plus a mini-grid you drag across to sketch any rectangle. Keys or click apply; arrows move the highlight with a live preview outline on the screen; Tab retargets another display; Esc closes.
• Saved layouts (in the palette): press S to snapshot every window's position; chips A/B/C restore the three most recent — "meeting mode" / "deep-work mode" in one keystroke. Restore repositions windows of running apps (matched by app + title); it never launches anything.
• hold + Return → maximize.
• ⌘Tab, fixed (no hotkey needed) → works like Windows Alt-Tab: window-level most-recently-used switching with a visual strip — hold ⌘ and rows of live window thumbnails appear (most recent first, then every other open window, up to 24), Tab or arrow keys move the highlight (←/→ step, ↑/↓ jump rows), ⇧⌘Tab moves backwards, release ⌘ to switch, Esc to cancel. Two Excel windows are two entries, unlike macOS's app-level switcher. Replaces the macOS app switcher while enabled — toggle in Settings ▸ Windows. ⌘` keeps its stock same-app cycling.
• Right after a snap, Snap Assist offers the other windows — click one to fill the empty space (Windows-style).
• hold + 3 → grid draw mode: drag across an on-screen grid to place the window (Moom-style). Grid dimensions are configurable.

Clipboard & cursor

• hold + P → Advanced Paste: a palette to paste the clipboard as plain text or transform it — clean text, single line, case changes, numbers only, contacts one per line (an Outlook/Gmail recipient line like Doug Lepping <[email protected]>; Willie Meredith <[email protected]> pastes as name, address, blank line, per person), then summarize, rewrite, bullets, markdown, or translate (AI transforms use your configured cloud model).

• hold + H → Clipboard history (the Windows Win+V gap): a drawer slides in from the right edge with your recent copies — text AND images (screenshots included), with thumbnails; scroll for older ones. Digits, arrows or a click pick one (Esc, hold + H again, or a click anywhere else slides it away), it pastes into the app you came from and becomes the current clipboard. Recording is local (SQLite; last 200 text clips + 25 images, images capped at 5MB each), and anything marked concealed/transient by password managers is never recorded. Toggle in Settings ▸ General.

• hold + D → New Document (the Windows "New ▸" gap): a small menu of document types — Word, Excel, Text, RTF, Markdown — creates the chosen blank file in the current Finder folder and selects it for renaming. Needs a one-time Automation permission for Finder (to read which folder is open); the Office files are valid minimal OOXML.

• hold + K → Color Picker (the PowerToys Color Picker gap): the system magnifier loupe appears — hover any pixel anywhere on screen and click. A little palette then shows that color in every common format — HEX · RGB · HSL · HSV · CMYK, plus bare hex, raw values, and 0–1 floats — pick one (digit, ↑↓+↵, or click) and it's copied to your clipboard. Works across Retina/HDR displays; Esc cancels. No screen-recording grant needed — it's the macOS native sampler.

• hold + M → Find My Mouse: dims every screen and sweeps a ring onto the cursor.

• Notch strip — one live glance in the camera housing, shared by several features rather than owned by any one of them. Collapsed it is a row of dots flanking the notch: one per Claude Code / Codex / Cursor session, tinted by state, plus a quota dot once a usage window runs low. Hover a dot and the housing opens into the agent list — one row per session with its app's icon, state, model and program chips, branch, task, what it is doing right now, how long it has run and what it has spent; a waiting permission carries ✓ / ✕ on its row, so you answer without opening anything, and the last few finished sessions sit dimmed under , each with an ✕ to clear it. The band beside the camera says how many are active and carries ↻ and the module launcher. The module row stays with the list, so you can go straight from a session to Weather or Disk without folding it away. Click a dot to pin it open; — another app, another window — and it folds back into the housing (the ✕ in the band beside the camera does the same). Hover the at the end of the strip and a single row of module icons opens — the same row, same size, that a module’s tab bar shows, so nothing moves under the cursor when one opens (Usage, Hotkeys, Snap, Clock, a Calendar month with today ringed, a MacNotch-style Weather panel with feels-like, humidity, wind, UV, rain chance, pressure, sunrise/sunset and the next hours for each of your cities, a panel — CPU, GPU and memory gauges over load, processes, uptime, memory pressure, swap, and the battery’s charge, health, wattage and temperature — a mirror to check yourself before a call — it opens on a click, never a hover, and the camera runs only while the panel is open — a panel — each volume’s used/free bar with an ⏏ for anything ejectable, over the live read/write throughput and since-boot totals of each physical device — Ask, an Agent Pad tile that shows the agent list in the notch on hover, a Macro Pad tile that toggles the pad, and a ⚙ Settings tile that opens the app's settings window); move away and it folds back, click one to keep it, and a tab row moves between them or out. When it opens it grows out of the camera housing — the outer top corners round off and a large concave cove sweeps up into each side of the notch, so the panel reads as flowing out of the housing rather than a slab hung under it. Modules stay a small panel below the housing — the notch is a glance, not a window; the agent list is the one shape allowed to go wide — and with no agents running it keeps just the beside the camera, so the panels stay one hover away. Everything is optional in Settings ▸ Notch.

Quick Capture & Connections

• hold + N (or any letter you assign) → Quick Capture: a small input box pops up — type or dictate a line, press ⏎, and it's POSTed to a connection you configure: a todo app, an n8n webhook, any HTTP endpoint. Add as many connections as you like in Settings ▸ Connections, each on its own leader letter, each with its own endpoint, auth header + token (stored in a private owner-only file), and JSON body template.
• Inline fields, parsed as you type: call Rhett tomorrow p1 @calls → title "call Rhett" with priority 1, tomorrow's due date, and context "calls" — the panel shows what it parsed before you send. Natural-language dates ("friday", "july 12") via Apple's on-device date detector.
• Template placeholders: %TEXT% (the captured line, JSON-escaped), %TODAY% (today as a Unix-seconds due date), %PRIORITY% (p0–p3, defaults 0), %DUE% (yyyy-mm-dd), %DUE_TS% (parsed due date in Unix seconds, today when absent), %CONTEXT% (@word). Example: {"title":"%TEXT%","priority":%PRIORITY%,"duedate":%DUE_TS%,"context":"%CONTEXT%"}.
• If a POST fails, the panel re-opens pre-filled so the text is never lost.

Chord mode (multi-letter shortcuts, in the spirit of KeyChord)

• hold + Space → chord mode. Now let go of the hotkey and keep just Space held while you type a short code; let go of Space → it runs. (Keeping the hotkey held and tapping Space works too — the code runs when the last of Space and the hotkey comes up.) Before you type, the panel shows everything — your chords and every menu item, grouped by menu and flowed into columns across the screen for big apps like Excel — and narrows as you type. A panel down the left edge lists every code for the app in front and narrows as you type; the typed part is dimmed, the letters still to type are in the accent, and an exact match is highlighted. ⌫ backs up a letter, Esc leaves without running anything. Nothing you type in chord mode ever reaches the app.
• Type a name instead of a code: two or more letters also match item names — filt finds AutoFilter, Advanced Filter… and Clear Filters in Excel. Name matches get their own section, listed first when no code matches exactly; the highlighted row is what letting go runs. An exact code always wins.
• Type a menu's name to browse it: view lists everything in the View menu (letting go on just the name runs nothing); keep typing to narrow inside it — viewfo → Formula Bar, Formula Builder, Header and Footer… — and letting go runs the top one. Three letters of a menu name (vie) bring that menu up too. A menu name also reaches the same-named ribbon tab (below): viewgr → Gridlines.
• Office ribbon (Excel, Word, PowerPoint, Outlook), Windows-Alt style: every tab has a dot code — .h Home, .i Insert, .v View, .da Data — and each command its tab's code plus its initials: .vgr Gridlines, .vzs Zoom to Selection. Typing a tab's code or name opens that tab in the app (the way Alt+W opens View) and lists its commands; letting go on a command presses it, and the ribbon goes back to the tab you were on (it stays put for a dropdown, and Esc puts it back too). Accessibility only shows the selected tab's buttons, so every tab chord mode opens is remembered (ribbon-cache.json) and its commands are then found by name from anywhere — gridl → Gridlines. .. learns the whole ribbon at once: it opens every tab (~3 s, the ribbon visibly flicks) and goes back. Toggle in Settings ▸ Keys ▸ Chord mode.
• Every menu item: no per-menu limit — Excel's Insert menu is listed in full.
• Menu chords: every menu item of the front app gets a code from its title's initials — → , → , → (one-word titles take two letters; joined-up words split at the capital, so → ). No setup, every app. Clashes are settled in reading order: the first item keeps the code, later ones add the next letter of their last word ( → ). The same menus always give the same codes, greyed-out items keep theirs (shown dimmed), and a quoted selection in a title () never changes a code. The menus are read fresh each time — tens of milliseconds — so what you see is the menu as it is now.

T / R / S / G / F need Screen Recording permission the first time (System Settings ▸ Privacy & Security ▸ Screen Recording, then quit & reopen). OCR is also on the menu-bar mic ▸ Capture Text from Screen.

Cloud cleanup (optional, opt-in)

By default the AI cleanup runs on-device (free, private). If you want a frontier model for smarter rewriting, set Cleanup to Claude or OpenAI in Settings and paste an API key (stored in your macOS Keychain). Only the transcribed text is sent to the provider — never your audio; transcription always stays on-device. Any cloud failure/timeout falls back to the deterministic local cleanup. The default Claude model is claude-haiku-4-5 (fast); switch to claude-opus-4-8 in Settings for higher quality.

Privacy

No network code exists in this repository. Speech models and the LLM are system-managed Apple on-device assets. Verify: grep -ri "http" Sources/ finds nothing but this README.