Shipped, one careful build at a time.

Crooked scans are finally straight — the deskew engine was rebuilt end to end. It had been measuring skew in a distorted coordinate space, over-rotating portrait pages by ~29% and nudging near-straight pages visibly crooked. Angles are now measured correctly, badly skewed pages (up to 30°) get corrected instead of skipped, rotation never clips ink at the page edge, and the page is cropped back to the paper afterward. Validated against a corpus of real scanner output.

Long paper support — Long paper (receipts, legal) in Settings → Scanner asks the scanner for its longest page instead of stopping at letter length. The scanner detects the end of the sheet; if it can't, ShabuBox safely falls back to normal length. Long pages also pass through image cleanup without being cropped, rotated sideways, or dropped as blank.

Whitened edge bands removed — pages from white-backed sheet feeders no longer keep a strip of scanner backing on the edge, even when the backing is nearly the same color as the paper.

Reprocess with feedback — Reprocess scan now remembers how a page was originally scanned and treats it the same way, and every view shows a live “Processing…” indicator while it runs.

Favorites, notes, and custom fields — star documents, attach free-form notes, and define your own typed fields, including a currency type with proper money formatting.

Typed tables — documents of the same type (statements, receipts, invoices…) can be viewed as a sortable table of their fields, with column sorting.

Inspector redesign — a cleaner document inspector, unified collections, and faster search.

Trustworthy file identity — the content-hash identity behind deduplication and thumbnails was hardened end to end: thumbnails can no longer show the wrong document’s content, rescans update atomically, and a save race can no longer lose a thumbnail.

Solid-black scans with Auto-levels — pages with dense black text or off-white paper could collapse the contrast stretch and render entirely black. The stretch is now mathematically incapable of failing this way, and mid-tones come out more faithful everywhere.

Background cleanup darkened pages — the whitening filter had its color adjustment inverted, darkening every page it touched instead of lifting the paper color to white. Gray and tinted paper now cleans up to a proper white background with text left dark.

Confidence gating — GLiNER extractions now carry confidence scores, and low-confidence guesses are skipped instead of cluttering your entity list.

Reprocess Entities in Settings → Maintenance re-runs extraction on every document in the library, so accuracy improvements apply retroactively.

Noise filter round two catches kinship terms, document roles, and structure labels that slipped through the first pass. Concurrent reprocess and typing passes no longer collide.

Scan image processing pipeline — a multi-step chain cleans up scanned pages before OCR, improving text extraction on noisy or skewed originals.

Actionable USB error messages — when a scan fails, the alert tells you what went wrong and what to try, instead of a generic error.

Drag to refile — drop files onto sidebar folders to move them where they belong.

Repair Library now surfaces pruned tombstones in its summary. The postal-address fold self-heals when its bookkeeping outlives the table.

Map section — every address extracted from your documents, geocoded offline from a bundled Census gazetteer. Three-tier collapse by state and city, an evidence-picked filter to narrow by person, and tap any pin to jump to the source document.

Addresses as a fourth entity kind alongside people, organizations, and places — browsable from the Entities sidebar with the same list-and-detail pattern.

Address normalizer — standardizes and deterministically groups address variants so "123 Main St" and "123 Main Street, Apt 2" cluster together.

Light, Dark, or Follow the Mac — choose in Settings. Every design token adapts, from accents through elevated surfaces, backed by an appearance-parameterized snapshot harness.

On-device document summaries — per-collection, with LLM summaries and larger thumbnails. Sensitive documents route through the local model; clean ones use Apple Intelligence.

Notes section with inline add, and a Timeline toggle between added date and document date.

Summaries now persist across navigation instead of re-summarizing from zero.

Trash and Duplicates sections in the sidebar, with bulk Restore and Delete Permanently.

Re-adding a deleted file restores it. Interrupted deletes finish. Tombstones can't be filed out of the trash. Restored files are fully re-indexed.

Help → Report an Issue — in-app issue filing via GitTickets, with a privacy-aware relay. My Reports lets you check on them without leaving the app.

"Most capable local model" ceiling — opt into a larger chat model, gated by available RAM, with explicit consent before any large download.

Scanning, ingestion, and entity extraction are more resilient: timeboxed OCR, real download timeouts, and a stalled inbox row never locks the user out.

Menu bar icons, persistent page rotation, entity suppress/hide, move/refile from Folder Health, and app icon v3.

App Sandbox — Release builds now run sandboxed with iCloud entitlements and a full privacy manifest. A one-time migration moves the library into the container.

In-process SANE scanning — the ScanSnap iX1300 now scans inside the sandbox without ScanSnap Home. eSCL/AirScan declares Bonjour correctly.

App icon v3 — refreshed design.

Vision OCR is timeboxed so a wedged ML stack can't stall the pipeline. GLiNER downloads stay out of the GPU gate. A stalled "Processing..." row never locks the user out.

Scanned-document naming settles after extraction; the Cancel button no longer sticks.

Licensing uses the data-protection keychain — no more password prompt on launch.

Paddle checkout live on shabubox.com; download page flipped to the ready card.

Ask your documents. Put a question to a collection in plain language and get an answer with numbered citations back to the exact page. Apple Intelligence answers on-device; a local model or direct passages cover anything it declines — so it never dead-ends and never phones home.

Files itself, and knows when not to. File a few documents and ShabuBox learns your folders — auto-filing when it's confident, suggesting when it's close, abstaining when it isn't sure.

People, organizations, and places from every document become a browsable, cross-linked index.

Scan drives the Fujitsu ScanSnap iX1300 over USB, works with any eSCL/AirScan network scanner, or watches a folder — PDF or JPEG, duplex, colour or mono.

⌘K search across every OCR'd word, note, entity, collection, and reminder. Optional on-device semantic search finds documents by meaning too.

Markdown notes with wiki-links, a calendar built from the dates inside your documents, smart collections that keep themselves current, reminders that sync with the macOS Reminders app, and a Folder Health view that shows what needs attention.

Documents stay ordinary files in ordinary folders; the database is just a rebuildable index. Optional iCloud Drive sync keeps a library in step across your own Macs. Everything runs on your Mac — no account, no server, no telemetry.

Free for 30 days with every feature unlocked, then a one-time $24.99 — no subscription, nothing renews. ShabuBox now updates itself in place through Sparkle, each update signed and verified.

Requires macOS 26 or later on an Apple Silicon Mac.