Lumen Reader: a screen reader you can ask about the screen
A screen reader for macOS that speaks each control the way VoiceOver does, and answers questions about what is on screen. Its AI runs on the Mac, and it makes no network requests.
- Role
- Designed and built, with Claude as a coding agent
- Year
- 2026
- Built with
- Swift, macOS Accessibility API, Apple Foundation Models, Screen readers, On-device AI
Lumen Reader is a screen reader for macOS that I am building. It follows keyboard focus and speaks each control: what it is called, its state, and what kind of thing it is, as in "Subscribe, unchecked, checkbox." Its keys are VoiceOver's, so anyone who already uses a Mac this way keeps the keys they know.
On top of that, you can ask it about the screen. Ask "Where is the search field?" and it moves its cursor there. Ask "Is there a video?" and it answers, or lists every match for you to choose from. That runs on Apple's on-device model, so nothing you ask, say or show leaves the Mac. Lumen makes no network requests at all.
It started with a request for captions
Years ago I taught interaction design in Miami, and that is where I met Mark. He has been deaf since birth. When our meetups moved online, he asked for captions and a sign language interpreter.
He was right to ask, and his friendship changed how I work. Accessibility belongs at the start of a project, before anyone has to ask for it.
A screen reader is that idea at its most literal. It decides whether a blind person can use the computer at all, so I wanted to build one myself and learn what it asks of the software around it.
It fills the gaps it finds
Most of what goes wrong for a screen reader user is somebody else's missing label. Lumen tries to close some of those gaps itself, and says when it has.
A button with no name gets a name guessed from the text around it, and Lumen says it is a guess: "Send message, guessed name." Before you send a form, it can check for required fields left empty, typos, and an attachment the message mentions but does not carry. It can describe a picture or a chart, read a table back in plain language, and read just a page's main content, without the menus and banners around it.
Everything it says also appears in a caption panel, with the current word highlighted. Speech is never the only way to get it.
Built with an agent, like the rest of my work
I am building Lumen with Claude as a coding agent, and every commit in its repository was written that way. It is the same way of working I describe in the agent tooling work: a person decides what it should do and checks what comes back.
It runs on macOS 14 or later. The features that ask the on-device model need macOS 26 with Apple Intelligence turned on.
What I can't claim yet
There is no public release yet, so nobody can install it from the website today. I can tell you what it does on my Mac. Whether it works for the people who rely on a screen reader every day is something only their use of it will show, and that comes after it ships.
If you use a screen reader on a Mac and want to try it, the Lumen Reader website (opens in a new tab) is where it will be.
Or press ←the left arrow and →the right arrow.