Ask about your screen
Reflick sees what you see, answers out loud, and can fly over and draw on the thing it means.
Reflick lives in your menu bar and your notch. Hold two keys and ask about the screen, or say “reflick agent” and it does the whole task itself: real clicks, real keystrokes, any app.
Reflick lives in the menu bar and the notch. There is nothing to open and nothing to switch to. You hold a key and speak, wherever you already are.
Reflick sees what you see, answers out loud, and can fly over and draw on the thing it means.
Types what you say. Or say “reply that Tuesday works” and it reads the screen and writes the reply itself.
Draw around a chart, a molecule, one paragraph. The traced shape goes to the model, not a bounding box around it.
Real clicks, real keystrokes, across any app. One agent run however many steps it takes.
Seven rungs, cheapest first. Five of them never touch the network. They read the Accessibility tree, the system, or a shortcut Reflick already knows. This ordering is the product, not an optimisation.
Read from the Accessibility tree, in the app developer’s own words. More accurate than a vision model, which is inferring from a picture of an icon.
“Hello”, “thanks”, “yes”, “stop”. Answered from a table.
Battery, storage, Wi‑Fi, volume, brightness, the time. Read from the system.
83 skills across 17 apps. The shortcut-backed ones execute with no model call at all.
Mail, Notes, Calendar, Files, Browser, Clipboard, scripted directly. No Full Disk Access required.
“Where is the export button.” Routed to the small model. Talking is always free.
Reasoning, summarising, writing, walkthroughs, and the agent loop. Only an agent task costs a run.
Each one runs locally through scripting or the Accessibility API. Type below to find one.
A skill is one thing Reflick knows how to do in one app. The ones backed by a keyboard shortcut run instantly, with no model call. Skills can be shared as files, and a shared skill is always stripped of its shortcut, because somebody handed a file cannot check what a keystroke would press.
It opens the browser, runs the query, reads the results back and leaves the map on screen with the nearest options ranked. You never touch the keyboard.
It applies the filters, reads every listing on the page, and comes back with a number instead of a link. This is the part a chat window cannot do.
Ask what's in front of you and Reflick reads your actual display through the Accessibility tree: which windows are open and what is in them. Then it answers out loud.
It opens Photo Booth, captures the frame and describes what's physically in front of you, then uses that as the input for whatever it does next.
It looked at your wall, worked out what the painting was, wrote its own search query from that, and left the references open in a tab. Then it closed the ones it didn't need.
It opens the app, finds what you asked for, presses play, and tells you it's running. Then it tidies up the windows it opened along the way.
All on your Mac, all readable as plain markdown you can open in Finder, all erasable in one click from Settings › Privacy.
Three times on the same weekday within the same hour and Reflick says so once, in a panel you opened: “you usually ask me to do this around now.” It offers a sentence, never a button that runs something.
Speech goes in one end. Native macOS calls come out the other. Nothing in between is hidden from you.
Deepgram nova-3Streaming websocket, with AssemblyAI, OpenAI and Apple Speech as fallbacksCartesia sonic-3Streams. First audio in ~0.4s regardless of answer lengthgpt-5.6 / gpt-5.6-lunaFrontier by default, small model for simple screen questionsgpt-realtimeCan see the screen mid-conversationgpt-5.6-terraComputer use, driving real CGEventsapi.reflick.myFly.io. Holds every API key, so nothing sensitive ships in the appReflick reads the screen you are already looking at. It does not ask for Full Disk Access, and it does not read your message databases.
Accessibility, Screen Recording, Microphone. Granted once in System Settings, revocable any time.
Never requested. Your Mail, Messages and Photos databases are not read.
The native capability layer is covered by an automated suite, not by hoping the model gets it right.
An agent run is spent when Reflick does something for you, once per task however many steps it takes. Talking is free and unlimited on every plan including Free, so asking about your screen, pointing at things and dictating never cost a run.
Refilled every month. Talking, pointing and dictating stay free and unlimited on top of that.
Download200 agent runs a month, for people who hand Reflick the whole task instead of asking about the screen.
Get Plus900 agent runs a month. For the day it stops being a tool you reach for and becomes how you work.
Get StudioOr €39 for 400. No subscription and no expiry. Purchased runs never run out at the end of a month.
Buy a packAnnual billing is two months free on Plus and Studio. Free runs refresh monthly and do not accumulate; a month spends its free allowance first, so a pack bought in January is still there in March.
Reflick is in final testing. It ships as a signed, notarised app: grant three permissions, hold ⌃⌥ and say something. That’s the whole setup.