Scanwise

Add your paperwork — phone photos, scans, PDFs, Office files, a whole folder as a .zip, files from Google Drive or OneDrive, or pasted text — and Scanwise tells you what it says and what to do about it.

1. Extract. Each file becomes a source (S1, S2…). Its text is read in your browser, with a confidence score for every word (green, amber, red).
2. Understand. Gemini reads the ticked sources together and writes one summary: key factors, key points, open questions and caveats, each marked with the source it came from, like [S1].
3. Act. Solutions are marked in the sources or suggested. Ask follow-up questions in a chat grounded in the sources — by typing, by voice, or by selecting a passage and choosing Quote in chat. Answers cite their sources too; click a citation to open it. Export everything as Markdown.

No account is needed to read files: Upload a document on the landing page reads them on your device. The summary, chat, AI vision and voice transcription need a Google sign-in; whatever was read before signing in moves into the account. History is kept per Google account in the browser (IndexedDB); before sign-in it lasts only for the tab. There is no server-side storage.

Layout

A conversation works like a notebook. On the left, Sources lists every source with a tick box: only ticked sources go to the AI, and when the set changes the summary offers Update summary. Clicking a source opens its preview and text, where you can fix misread words or quote a passage. On the right, the summary sits above the chat. On phones the two columns become tabs.

Add sources offers Upload files (several at once, or a .zip — each document inside becomes a source), Google Drive, OneDrive and Paste text. Drive and OneDrive appear only when their settings are present. Files are read one after another, each with its own progress bar. A conversation holds up to 25 sources.

What it reads

Family	Formats	How
Images	PNG, JPEG, WebP, HEIC/HEIF, TIFF (every page), GIF, BMP, AVIF, SVG	On-device OCR (tesseract.js), or Gemini vision
PDF	Digital and scanned	pdf.js text layer; scanned pages are OCR'd (up to 40)
Office & OpenDocument	DOCX, PPTX, XLSX, ODT, ODP, ODS, EPUB, RTF	Unzipped and parsed in the browser
Text & web	TXT, Markdown, CSV, JSON, HTML, XML, YAML	Read directly
Archives	ZIP	Unpacked in the browser; each file inside is read as above
Google Drive	Google Docs, Sheets, Slides and any file above	Docs, Sheets and Slides are exported as DOCX, XLSX and PPTX

Reading has three modes: Auto (on-device, with the AI re-reading only low-confidence images), On-device (the file never leaves the browser) and AI vision (best for handwriting and messy scans). Only the extracted text is sent for analysis, unless AI vision is chosen.

Voice input uses the browser's own speech recognition (Chrome, Edge, Safari). Other browsers record and send the audio to Gemini for transcription, which needs a sign-in.

Stack

React 19 + TypeScript + Vite 8, Vercel functions in api/ (Web-standard handlers), Google Identity Services sign-in with an HttpOnly JWT session cookie (jose), Gemini via @google/genai, tesseract.js, pdf.js, fflate, utif2, heic2any, and @azure/msal-browser (loaded only when someone opens OneDrive).

Run locally

npm install
cp .env.example .env.local   # then fill in the values; the file explains where each one comes from
npm run dev                  # http://localhost:3000 — the /api functions run inside the Vite dev server

With no .env.local at all, reading files still works (Upload a document); sign-in and the AI stay off until GOOGLE_CLIENT_ID and GEMINI_API_KEY are set. SESSION_SECRET is optional in dev and required in production. Google Drive needs GOOGLE_API_KEY and GOOGLE_APP_ID as well; OneDrive needs MICROSOFT_CLIENT_ID.

Checks

npm run typecheck            # browser, server and config TypeScript projects
npm test                     # unit tests (vitest)
npm run lint                 # oxlint
npm run build && npm run test:e2e   # real browser run of the production build

The end-to-end run mocks the /api routes (signed-in user, analysis, chat), so it needs no keys. It reads every file type, several files at once, a .zip and pasted text; checks citations, quoting, the summary going out of date, the signed-out path and phone layouts. On-device OCR downloads its engine and language data from jsDelivr the first time, so it does need a network connection. The Drive and OneDrive pickers need real accounts and are not covered by it.

Deploy (Vercel)

Import the repository; Vite is detected and vercel.json sets the function timeouts and security headers. Add GOOGLE_CLIENT_ID, GEMINI_API_KEY and SESSION_SECRET (optionally GEMINI_MODEL, GOOGLE_API_KEY, GOOGLE_APP_ID and MICROSOFT_CLIENT_ID) under Project → Settings → Environment Variables, and add the production URL to the OAuth client's Authorized JavaScript origins. For OneDrive, also add https://<your-domain>/auth/microsoft.html as a Single-page application redirect URI in the Azure app registration.