VaultSnip privacy policy
Short version: VaultSnip stores nothing about you or your images.
- No accounts, no backend, no database. VaultSnip is static code that runs in your browser.
- Your images stay on your device. Text recognition (OCR), masking and the safety check run locally in your browser. Images and results are held in memory and disappear when you close the tab.
- Nothing leaves your device. Reading (OCR), masking, the safety check and the replica all run in your browser, in this tab. VaultSnip sends your screenshot, the masked image, the text it reads and the synthetic text nowhere: there is no AI service, no API key and no upload. The page's security policy only lets it contact VaultSnip's own site, where its model files come from.
- Numbers in the masked image are invented. By default each number (amounts, percentages, scores, temperatures, counts) is replaced with a different, made-up value of the same size: $4.2M might show as $5.6M. The real values never leave your device, but the rough size of each number does (same number of digits, same unit). Names, labels, emails and IDs stay covered by grey boxes, and the image carries a note saying its numbers are made up. Switch it off under Redaction policy to cover numbers with grey boxes instead.
- Customer comments (optional). If you tick Keep the wording of comments, the ordinary words of quotes and free-text feedback stay readable in the masked image, and can appear in the replica and its export. Names, companies, emails, numbers and IDs inside them stay masked, and the replica only quotes words that were readable in the image you approved. It is off by default.
- Samples use a bundled, fictional screenshot.
- The replica is your own screenshot, changed in place. VaultSnip wipes each hidden box with the background around it and draws synthetic text back in the same place, size and colour, in your browser; everything else in the picture stays as it was. The synthetic text is made on your device: numbers of the same size, made-up names and companies, emails only at the reserved example.com, example.org and example.net domains. Pointing at a replaced box shows the synthetic value, its kind and the visible words beside it, never the hidden text.
- What the replica and the masked image still show. Only text is replaced. Everything drawn stays as in your screenshot: the layout, the visible labels, and the shapes of the charts (bar lengths, line trends, slice sizes, map shading). Numbers are invented but keep their size (the same number of digits and unit), and axis labels are rebuilt on a new scale, yet a chart's proportions are still the real ones. If the shape of a chart is itself confidential, choose Blur a chart on the review screen and drag over it, or Blur all charts to have every chart found and blurred at once (found in your browser, from where the picture's colours are): that area is blurred in your browser, in the masked image and the replica, so its shape cannot be read.
- On-device name check (on by default). A small AI model looks for safe words used as names ("May Chen", "Delta Airlines") and adds masks for them. It runs in your browser, on this device; the text it reads never leaves it. The model is an app file (19 MB) served by VaultSnip itself, never by a third party; extension packages built for the stores include it. It only adds masks, and you can switch it off under Redaction policy.
- Unread-text check (on by default). A second small AI model finds every patch of text in the screenshot without reading it, so text the reader missed (tiny, rotated or faint labels) is masked too; a third reads such a patch once more, one line at a time, to tell a safe label from a name. They run in your browser, on this device, and see only the image's pixels; nothing leaves the device. The models are app files (5 MB and 7 MB) served by VaultSnip itself, never by a third party; extension packages built for the stores include them. It only adds masks, and you can switch it off under Redaction policy.
- What your browser keeps. The app's model files (the English OCR model and the on-device AI models), cached so they download once. The text reader and its English language file start loading from VaultSnip's own site when you point at the drop area or a sample, or drag a file over the page, so they are ready by the time you have picked a file; nothing about you or your file is sent. On the web app, a service worker (part of the app) also keeps the app's own files (the page, its scripts, the reader's program and the on-device AI runtime) in the browser's cache, so the second visit opens at once and works offline; it caches only files from VaultSnip's own site, and never a screenshot, a result or anything read from an image. If you click Save under Redaction policy, the policy's name, client names and safe words are kept in this browser's local storage so you can load them again; nothing else is stored, and Delete removes a saved policy. Choose a policy name that does not itself give a client away if others use this browser. No screenshot or result is stored.
- Desktop app. The desktop app (built from
desktop/, not offered on the site) is the same page in a window of its own. Its hotkey takes a screenshot of your primary screen when you press it and hands the picture to the page in memory; once the check passes, the page puts the masked image on your clipboard without a click. The screenshot and the masked image are never written to a file by the app. The page is served from the app itself: it contacts no site, not even VaultSnip's own, so nothing can be sent anywhere. The app keeps running in the tray when its window is closed, so the hotkey still works; Quit in the tray menu ends it. It reads the screen only when you press the hotkey, click Snip the screen or choose it in the tray menu. - Blur text in charts (extension). "Blur text in charts on this page" (right-click menu, or Alt+Shift+B) reads the charts drawn on that page, finds sensitive text on your device (the same OCR, rules and unread-text check as the app, with the same models packed in the extension), and blurs it in place. Nothing is sent or stored, and running it again removes the blur. It runs only on the tab where you choose it.
- Extension permissions. No access to any website until you act on a tab, and none afterwards.
activeTab: capture or blur the tab you are on, only when you click the icon, the menu item or the shortcut.scripting: put the chart blur into that tab when you ask for it.offscreen: a hidden extension page that reads the chart text on your device.contextMenus: the "Blur text in charts on this page" menu item. None of them shows a warning when you install. - PDF input. A PDF you open is drawn page by page in your browser by PDF.js, which is part of the app (served by VaultSnip itself, 1.7 MB, loaded the first time you open a PDF). The PDF is not uploaded.
- Batch masking. Every screenshot or PDF page in a batch goes through the same reading, masking and six-pass check as a single screenshot, on your device. Only images whose check passes are saved; the rest are held back for you to review. The ZIP is made in your browser. Receipts made in batch mode say that no person looked over the image before it was saved.
- Downloads (the masked PNG, the replica PNG, a batch ZIP) are made in your browser and saved to your computer; nothing is sent.
- Safe-to-share receipt. Receipt (PNG) and Receipt (PDF) save one page, made in your browser, that lists how many items of each kind were hidden (never the hidden text), the result of the check, the date and time, the VaultSnip version and the SHA-256 fingerprint of the masked image file. It shows a small copy of the masked image, the same picture you approved. Nothing is sent or stored.
- Clipboard. VaultSnip reads the clipboard only when you paste (Ctrl+V / ⌘V) or click Paste a screenshot, and uses only a picture found there; your browser may ask you to allow it. It writes to the clipboard only when you click Copy masked image, Copy image (the replica picture, synthetic text only) or Copy link. No clipboard permission is added to the extension.
- Exports. Download image and Copy image save or copy the replica picture, made in your browser, with a strip below it saying its text is synthetic. Nothing is sent or stored.
- Feedback is your choice. The Feedback button prepares a report (VaultSnip version, browser, the step you were on and any error message, never your screenshot or dashboard text). Nothing is sent until you choose to continue on GitHub, where it is posted under your GitHub account, or to email it from your own mail app. GitHub's or your email provider's terms apply.
- Developer benchmarks are not part of the app. The accuracy tools in
tools/run only when a developer starts them, or in this project's CI. They draw fictional dashboards from invented data on the developer's machine and download nothing; the app and extension never run them. - No analytics, cookies or tracking.
Last updated: 4 November 2026.
Contact: use Feedback in the app, open an issue or discussion on GitHub, or report security concerns privately as described in SECURITY.md.