Take a screenshot, press Ctrl+V here, and get the text. No saving it to your desktop, no hunting through a file picker, no deleting it afterwards — the image is read straight out of the clipboard and never touches your disk or any server. Screenshots are also the input OCR handles best, because the text was drawn by a computer rather than photographed: flat, evenly lit, consistently shaped, and with none of the compression damage that blurs letter edges in a JPEG. The one thing that genuinely limits accuracy is resolution — ordinary body text in a screenshot is only about a dozen pixels tall, which is near the engine's lower bound. Zoom the app before you capture and the result improves noticeably.
| System | Shortcut |
|---|---|
| Windows | Windows + Shift + S — selects a region |
| macOS | Cmd + Shift + 4 — selects a region |
| macOS, whole screen | Cmd + Shift + 3 |
| Linux (GNOME) | Shift + Print Screen |
| Then, here | Ctrl + V or Cmd + V |
Take it as you normally would — Print Screen or Windows+Shift+S on Windows, Cmd+Shift+4 on a Mac — then click this page and press Ctrl+V, or Cmd+V. It is read straight from the clipboard, so there is no file to save or delete.
No, and it is not written to disk either. Recognition happens in your browser, and the image exists only in memory for as long as the tab is open. Screenshots often catch more of the screen than you intended, which is a good reason for them not to leave your machine.
Because the text was rendered rather than photographed. There is no angle, no shadow, no focus problem, and no JPEG compression rounding off the letter edges. The recogniser is seeing exactly the pixels the application drew.
Increase the size before capturing — zoom the page or application, or capture on a high-density display. Small text is the main failure mode here, because there are simply not enough pixels per character for the engine to work with.
No. You get the text in reading order, but columns and cells are flattened into lines. Reconstructing a grid from a bitmap is a much harder problem than reading the characters.
Yes — drop a folder of images in rather than pasting, and they are processed as a batch.
The distinguishing feature here is the clipboard, not the recogniser. A screenshot exists in the clipboard the instant it is taken, and the usual workflow — save it to disk, find it in a file picker, upload it, delete it afterwards — is entirely avoidable friction for something that is going to be read once and discarded. Pasting reads the image straight from the clipboard event, so nothing is ever written to disk. Recognition itself is Tesseract compiled to WebAssembly, running on your device, and screen captures are the input it handles best: the text was rendered by a computer rather than photographed, so it is flat, evenly lit, consistently shaped and free of the compression artefacts that damage letter edges in a JPEG. The real constraint is pixel density. Body text in a screenshot from a standard display gives roughly twelve to sixteen pixels of character height, near the lower bound of comfortable recognition, so zooming the application before capturing — or capturing on a high-density screen — improves accuracy more than any setting on this page. Layout is flattened to reading order: a screenshot of a table returns its text in sequence rather than in cells, because reconstructing a grid from a bitmap is a different and much harder problem than reading the characters in it.