franzenzenhofer/big-arrow-on-the-screen

big-arrow-on-the-screen (bigarrow) is a macOS command-line tool, plus a skill for Claude Code and Codex, that draws an arrow and a sign on top of every window. Clicks go through, your keyboard focus stays put, and the arrow removes itself. MIT licensed.

Real apps · Copy & paste · What is it for? · Install · Commands · Looks · Creative arrows · Starred by · FAQ · How we know it works · For agents · Plan · Prior art · License

Your AI agent can refactor a monorepo, write a migration and explain monads, but when it needs you to click one button it prints "please click Allow in the dialog" into a terminal you are not looking at. bigarrow gives it a finger.

bigarrow point --element "Allow" --app "System Settings" --text "Franz, click Allow: Ghostty may control your Mac"

One transparent window above everything, on every display and every Space. Drawing needs no macOS permission at all. One Swift binary: no daemon, no menu-bar icon, no account, no telemetry, and, we checked twice, no AI inside. It is an arrow.

Real apps on a real Mac (macOS 27), the real bigarrow, staged by scripts/real-scenes.sh.

"The most annoying change in Mac update history", fixed in six arrows (MP4).

open "x-apple.systempreferences:com.apple.preference.security?Privacy_Accessibility"
bigarrow start --element Terminal_Toggle --app "System Settings" \
  --text "Franz, switch this on: Terminal may control your Mac" --from right --color green --close-button
bigarrow start --element Add --role button --app "System Settings" \
  --text "Not in the list? Plus. Then find it." --from bottom-right --style ring --color orange --shape zigzag --size S

bigarrow start --element Animate --app Keynote --role radiobutton --text "1. Click Animate" --from right \
  --style ring --color purple
bigarrow start --element "Add an Effect" --app Keynote --text "2. Add an Effect" --from right \
  --style box --corners sharp --color "#FF9F0A" --shape straight
bigarrow start --element Play --app Keynote --role button --text "3. Press play. Bask in the applause." \
  --from top --color teal --shape zigzag --size S

bigarrow start --element PDF --role button --app TextEdit \
  --text "Mom, click PDF, then Save as PDF" --from bottom --color pink --border white-black
bigarrow start --element Cancel --role button --app TextEdit \
  --text "Not this one, Mom" --from bottom-right --color black --size S

bigarrow start --app "Google Chrome:Sourdough" --element "Sourdough - Wikipedia" --role radiobutton \
  --text "It's this tab, not the other 13" --from top --shape zigzag --color "#5856D6" --size L

--app "App:tab title" raises the window and selects the tab first.

bigarrow start --element "icon view" --role radiobutton --app Finder \
  --text "Not this one" --from top-left --style ring --color black --size S
bigarrow start --element Group --role menubutton --app Finder \
  --text "No, the other grid icon. This one." --from top --style box --color green

The whole guide as a 6-page PDF, every screenshot an arrow an agent drew. Download the PDF or read it here.

{{value}} in the sign becomes a copy button. Click, paste.

bigarrow start --element "PIN code" --role textfield --app "Google Chrome:Acme sign-in" --from right \
  --color blue --border-color yellow --text-color yellow --text "Franz, copy & paste the PIN {{482913}} here"

Fair question. Arrows have existed since roughly the Paleolithic. What changed: agents now do real work on your Mac, and sooner or later hit a step only a human may do, or one the human wants to learn.

• "Click Allow." Permission prompts, OAuth consent, "Open with...?". The agent finds the button but must not, or cannot, press it. It can point.
• "Your turn." 2FA, CAPTCHAs, passkeys, payments, signatures. It points, you decide, it continues.
• "Paste this." A code, a URL, a command: on the sign with a copy button.
• "I need you, and you're making coffee." --say reads the sign aloud. Your Mac will literally call you back to your desk.
• "Show me how." Ask how to do something in Blender, and the agent points at each control in turn. A product tour, minus the product.
• Helping someone else. On a parent's Mac or over a screen share, pointing beats "the button at the top, no, the other top".
• Demos and docs. Highlight while recording, or render straight to a PNG with --png.
• Debugging coordinates. Point at your Accessibility or Peekaboo coordinates and look. --dry-run --json says where it would point.

Not a screen annotator, not a click bot, not a screenshot tool. It never clicks, types or captures. It only points. Deliberately.

brew install franzenzenhofer/tap/bigarrow
bigarrow install-skill          # teaches Claude Code (~/.claude/skills) and Codex (~/.agents/skills)

From source: swift build -c release (Xcode 16+, macOS 14+), binary at .build/release/bigarrow. The binary is pure Swift; scripts/ only records screenshots and runs tests.

bigarrow point --element "Allow" --app "System Settings" --text "Franz, click Allow"   # by label
bigarrow point --at 760,500 --text "Franz, click HERE"                                 # by coordinate
bigarrow start --window "Safari:Inbox" --text "This window" && bigarrow stop            # until stopped

Every arrow ends by itself. Nobody has to clean up after an agent that forgot:

Targets: --at X,Y, --rect X,Y,W,H, --mouse, --window App[:title], --element Label --app App, --peekaboo ID --snapshot see.json. Coordinates are global top-left logical points, as Accessibility and Peekaboo report them; --display N makes them relative to one display.

--app App[:window or tab title] (or --window) first brings that app, window or Chrome/Safari tab to the front, because pointing at a window hidden behind your terminal is a special kind of unhelpful. If another app covers the target later, the arrow hides until it is visible again. --no-raise opts out. bigarrow elements --app X lists what --element can match; bigarrow doctor shows permissions and displays.

Every command takes --json. Exit codes: 0 ok, 2 bad input, 3 target not found, 4 permission missing. Agents love exit codes. Humans tolerate them.

It is an arrow, so we spent an unreasonable amount of time on how it looks.

Six looks over white, macOS grey, dark, black, red and a busy web page:

• --shape bend|straight|zigzag|spiral: zigzag for when it is really urgent; spiral loops once around the sign, for when it must be impossible to miss
• --style arrow|ring|box (rings and boxes are outlines, the target stays visible), --size S|M|L, --corners round|sharp
• --color red|orange|yellow|green|teal|blue|purple|pink|black|white|#RRGGBB
• --border shadow|white-black|black: white border with drop shadow (default), a thin black edge instead of the shadow, or just a thin black outline
• --border-color, --text-color, --edge-color, --close-color, --close-x-color; left out, each picks a readable colour
• {{value}} in --text becomes a copy button; --close-button puts an X inside the sign's right end; --follow moves with a window or element; --until-click ends on a click on the target; --say speaks the sign
• Several arrows at once keep their signs out of each other's way

scripts/gallery.py renders every combination and zooms into every sign-to-shaft joint (junctions), because a seam at the joint was, apparently, unacceptable.

Fake dialogs, real bigarrow, clean CI runner (BACKDROP_ARGS=--cover scripts/funny-scenes.sh). The dialogs are fake. The feelings are real.

--color green, because some decisions are easy

--shape zigzag --color orange

--shape spiral: once around the sign, then to the button

--color purple

--close-button, because the human gets the last word

three starts, one button, zero ambiguity

--style box --corners sharp, plus a lesson about macOS permissions

Nearly 400 stars in the first two days, from people whose GitHub profiles list Apple, NVIDIA, AMD, SAP, Salesforce, ServiceNow, Booking.com, Mercedes-Benz, SUSE, Oxide Computer, Posit, CoreWeave, Weights & Biases, OpenRouter, Metabase, InstaDeep, Benchling, Stainless and Under Armour, plus Stanford, Johns Hopkins, KTH and Oak Ridge National Laboratory.

Permissions? · In front without focus? · Steals my typing? · Click through? · Displays and Spaces? · CPU? · Web pages? · Can it trick me? · Token cost? · Why not an annotation app? · Is it AI?

Drawing: no. Some ways of finding the target do:

macOS grants these to the app that started bigarrow (Terminal, iTerm2, Ghostty, VS Code, Claude), never to bigarrow itself, so switch on that app. bigarrow doctor names it; a missing permission exits with code 4 and names the app and the settings pane.

On top and focused are two different things on macOS. The arrow sits at screen-saver level, above windows, dialogs and full-screen apps, but never becomes the active window.

No. That was the hardest bug in the project: NSApplication.run() quietly activates a process without a terminal. bigarrow pumps events itself, and the tests check that the frontmost app never changes.

Yes, everywhere except the sign and the shaft. A click there removes the arrow (it dims under the pointer to say so). Clicks on the target or near the head go straight to the app, without taking the focus.

Yes, yes, yes, yes. Negative coordinates included. Unplug a display while an arrow is on it and the arrow politely leaves. See the verification matrix.

1.4 % on a CI runner. Core Animation does the work in the render server.

In Electron apps, yes. In Chrome, only with help. Chrome needs --force-renderer-accessibility (or VoiceOver on); it ignores the usual request (verified October 2026). Chrome's own toolbar always works. Otherwise point at page coordinates, which the skill explains.

Say, by covering the Decline button? Not beyond what it can already do. An agent that runs shell commands as you can do far worse, so bigarrow gives it nothing new. Still, each checked by a test: boxes and rings are outlines, so the target stays visible; the sign keeps clear of the target (or overlaps it as little as possible); a click on sign or shaft removes the arrow; every arrow ends by itself. And the skill makes the sign say what your click does.

182 tokens, most of the time. That is the skill's description, the only part the agent always sees. The instructions, 1,528 tokens (Anthropic's token-count API, Claude Opus 5.5), load only when it decides to point; pane ids and look flags (1,118 more) only when it needs them.

Those are for humans drawing on screens. This is for programs pointing at things, from a shell, with exit codes. Twenty-six tools were checked first (research). None did this.

No. It is the least intelligent part of your AI stack, and proud of it.

• 104 automated tests: geometry, placement, joints, a golden image, copy buttons, recorded window-server, Accessibility and Peekaboo fixtures, and the real window server (click-through, focus never moves, stop timing). CI runs macOS 15; also passed on macOS 26 and 27.
• 18 behaviour checks on a clean runner (visual.yml): clicks on the X, a copy button (the clipboard holds the value, the arrow stays), --until-click, --follow, raising and --no-raise, hiding while covered, Chrome tab selection, owner exit, stop --hook, --say, full-screen, Stage Manager, Spaces, second and 2x displays, unplugging mid-arrow, CPU.
• A fresh agent given only the skill and "show Franz where the Reload button in Chrome is" found it by label and built the right command (transcript). It also found a bug, which is now a test.

skill/big-arrow/ (Agent Skills format, plus agents/openai.yaml for Codex) teaches the agent when to point, how to pick a target, to write a full sentence on the sign, to --say it when you are away, and to stop once you acted.

docs/plan/PLAN.md, docs/plan/TICKETS.md (generated from tickets.json), docs/decisions/, docs/research/, docs/verification/, docs/skill-tests/, CHANGELOG.md.

Copy and check icons: Feather (MIT, notice). Peekaboo (https://github.com/openclaw/Peekaboo) and Nameplate (https://github.com/steipete/Nameplate) by Peter Steinberger showed the overlay recipe and the skill packaging. Neither points with a labelled arrow. bigarrow reads Peekaboo's see --json as a target source.

MIT. Point responsibly.