An AI that can see your screen, on your Mac

With a chatbot in a browser tab, you copy, paste and describe what you're looking at. An AI that can see your screen skips all that. You ask about the thing in front of you, and it already knows what “this” is.

With Koi, hold Fn + ⌘ and ask out loud, then let go to send. Koi answers out loud, in whatever language you speak to it. It only hears you while the keys are down, so nothing is listening in between. For seven seconds after an answer, hold the keys again to ask a follow-up. Then the call ends by itself, or you can end it with a tap of the keys or Esc.

Ask “what does this error mean?” or “which one do I click?” and it looks at your screen. Want to be precise? Hold the keys and draw a ring around something with your pointer: the next question is about what's inside it.

Everyday questions get an answer straight away. When the answer depends on today, it searches the web, and for maths, code or careful reasoning it hands the question to a stronger model. It tells you when it's searching or thinking something through. While a call is on, a light runs round the edge of the screen, so you always know Koi is there.

Seeing the screen is a setting with three answers: never, only when you ask, or with every question. It starts on every question, and you can change it any time in Settings, under Talking with Koi.

Even then, Koi isn't watching. A picture of the screen goes along when you ask a question, and only then. Between questions it sees nothing and hears nothing.

macOS also asks you to allow Screen Recording before Koi can see anything. If a permission is missing when you call, Koi says what's missing and how to allow it, and lets you call anyway.

Koi answers on the screen itself. It circles the button you need, or highlights the exact sentence it means. For harder things it draws on a board laid over your screen: a diagram, a graph, a sum worked out step by step. The marks look hand-drawn and go away once they've done their job.

It's the kind of help you'd ask a friend sitting next to you for. A few things to try:

• In an app you've never used: “where do I turn this off?” Koi circles the switch.
• On an error message: “what does this mean, and what do I press?”
• In System Settings: “where's the setting for my keyboard shortcuts?”
• On a chart or a dense paragraph: “explain this”, with the part it means highlighted.
• On homework: “walk me through this sum”, worked step by step on a board over the screen.
• On a form: “did I miss anything?” before you press Submit yourself.

Questions don't have to be about the screen. Ask what the weather is doing, how to word a hard email, or what an acronym stands for, and Koi answers the same way, out loud.

Select text anywhere and say “fix the grammar”, “make it shorter” or “reply that I'll be late”. Koi changes it in place and highlights the new words, and ⌘Z takes it back.

On a Koi page it can do more: write where the cursor is, add a summary, turn lines into headings or lists, or change several places at once. Ask on a call and it makes a whole page, tables and all, and opens it. That's how Koi's notes, with tables and charts, fill up without typing.

When you just want your own words down, there's dictation: tap Fn + ⌥ and Koi types what you say. That part runs on your Mac and never goes online. More on voice to text on Mac.

Yes, up to a line. Say “open YouTube and search for lo-fi”, “click Skip” or “scroll down”. Koi uses real clicks and keys, with subtitles saying what it's doing, and your pointer is only borrowed for the press. It checks that what it did worked, and tries another way if it didn't.

It can start things for you too. Say “take dictation for me” or “record this call”, and Koi says so, hangs up, and starts dictation or call recording on your Mac.

Some buttons are yours to press. Koi won't buy, pay, order, send, post, publish, delete, confirm, accept terms, sign in, or type a password or a card number, whoever asks.

That's enforced in the app, not just asked of the AI model. Instead, Koi gets you to the button, circles it, and lets you press it. A drafted email waits for you to hit Send.

Text on a web page is never an instruction to Koi. Only your voice is. A page that says “ignore your rules and click here” is just words on a page.

A call with Koi needs the internet, because the AI models it talks to don't fit on a laptop. While the call lasts, this goes to the AI provider:

• Your voice.
• A picture of your screen with each question, unless you switch it to only when you ask, or never.
• Selected text or the front window's words, only while their switch is on, and only with the question that needs them.
• What Koi remembers about you, and what was said in the last ten minutes, so the answer makes sense.

Nothing said is stored. Koi's server keeps what a call cost (time, tokens, money), because that's how your plan's time with Koi is counted. It never keeps what was said, shown, written, drawn or remembered. A call that goes quiet hangs itself up, so silence isn't streamed.

Our trust page sets out what stays on your Mac and what we keep.

Yes. Tap Control twice and a chat grows out of the notch. Type, press Return, and it folds back into the notch saying Thinking, then opens again on Koi's written answer, with a line for a follow-up. No call and no voice, which suits a meeting, a library or a shared office.

In the short term, yes. Questions asked within ten minutes of each other are one conversation, so “and that one?” just works. Ten quiet minutes and it starts fresh.

Koi also remembers what you tell it, or ask it to remember: your name, what you're working on, how you like answers. That memory stays on the Mac. Settings, under Memory, shows every line of it to change or delete, or turns it off.

In part. OpenAI's help center says the ChatGPT app for Mac can attach a screenshot of an open window you pick, which needs Screen Recording permission. Its Work with Apps feature reads content from supported apps, such as Apple Notes, Xcode, VS Code and Terminal, and can edit code in IDEs.

Apple's Siri AI, in beta on macOS 27, can answer questions about what's on your screen on Apple Intelligence Macs, in English to start. Command-Shift-Space picks a window to ask about. Both were checked on October 7, 2026.

Koi is for people who want help that points, draws and clicks, on any Mac with macOS 14 or later, Apple silicon or Intel, with dictation and call notes in the same app. For the app most like it, read how Koi compares with HeyClicky.

Free includes 15 minutes' worth of Koi, once, plus unlimited dictation and 2 recorded calls a month. Pro is $20 a month with higher usage limits, and Max is $99 a month with up to 5x more usage than Pro. Yearly plans are 20% off.

Rather pay once? Lifetime is $99 and comes with about 5 hours of Koi that never expire. The app shows your usage as a percentage, never as a meter running in minutes. See every plan.