The words are
in there.

No. There is no upload route in the code. The only download is the language model, from this site, on a language's first use — after that, reading makes no new downloads at all.

Yes — that is the main case. Each page is rendered to an image in your browser and read one at a time, up to ten pages per run, so a long scan is read in parts. The page count and any skipped pages are shown before you start.

Yes. When a page reads poorly, CapyRead turns it in quarter-turns until it reads, estimates a slight skew from the words' own baselines and straightens it, then says what it did — "page 1 read 90°" — in the file card. Straight pages pay for none of this: one read, as before.

Usually not well. Tesseract is strong on printed text and weak on cursive, and the confidence badges will say so: paragraphs it is unsure about show "unsure" rather than pretending.

English plus seventeen more — German, French, Spanish, Italian, Portuguese, Dutch, Russian, Polish, Ukrainian, Turkish, Arabic, Hindi, Vietnamese, simplified and traditional Chinese, Japanese and Korean. Each language's model downloads once, from this site, when you first pick it.

That is a word the engine could not read with any confidence. CapyRead prints the gap instead of a wrong guess, so you can check that spot in the original and type over it in the editor.

The language model only — kept so a second run starts instantly, with a one-click "forget" control under The file. Your documents are never stored; close the tab and the suite forgets you.