Screenshots and OCR
Region capture with on-device text recognition, so a screenshot is as searchable as a note.
Cmd/Ctrl+Shift+Alt+M captures a region or window and OCRs it in the background, so the
recognized text joins full-text search -- a screenshot becomes findable by whatever it contains,
without anyone typing a caption.
macOS
Region selection shells out to screencapture -i -- the same interactive picker behind
Cmd+Shift+4, with window highlighting and marquee selection for free. Text recognition runs
through the Vision framework (VNRecognizeTextRequest), on-device, no network call and no
third-party dependency.
Vision reports one observation per text region it detects, not necessarily one per visual line -- a row with a wide gap in it (spaced-out terminal or table output) can come back as several separate observations for what a person would call one line. magpie groups observations by their bounding box (vertical center within half a text-height counts as the same line) before joining them, rather than trusting Vision's enumeration order to match reading order.
Linux
Region selection goes through the org.freedesktop.portal.Screenshot portal interface
(interactive: true) -- the standardized mechanism across X11 and Wayland, GNOME and KDE alike,
rather than a hand-rolled X11 selection overlay. Text recognition shells out to tesseract
if it's present on PATH.
Unverified on real hardware
The Linux path type-checks and passes its unit tests in CI on a real Ubuntu runner, but nobody has run the interactive portal flow on an actual Linux desktop session -- CI has no display or portal daemon to test that part against. Treat it as implemented-to-spec, not yet hands-on-verified. See Architecture.
Both platforms report capability honestly
capabilities() tells the UI whether screenshot capture and OCR are actually available right
now -- on Linux, that depends on whether a portal implementing the Screenshot interface is
reachable and whether tesseract is installed, neither of which can be assumed. The app adapts
rather than showing controls it can't back up.