SharifandClaude Opus 5
00991c1608
Search Console (Jun 4 – Sep 3) shows the page ranking on the first results page for "content credentials" queries while converting badly: "ai credentials remover" at 113 impressions and 0.88% CTR from position 6.7, "remove content credentials from image online" at 2.15% from position 5.0, "remove ai content credentials from image" at 0% from position 7.8. The title led with "EXIF metadata", a phrase that appears in none of the top queries, so the snippet never matched the intent behind the impressions. SEO - Title, description and H1 rebuilt around "remove content credentials from an image"; EXIF kept as a secondary term. - Added a how-to section and a nine-question FAQ answering the question queries that had no content behind them, including the LinkedIn AI label and the Instagram AI tag. - Structured data as an @graph: WebSite (alternateName "Untrace AI" for the unserved brand query), Organization, WebApplication, HowTo, FAQPage. - Fixed sitemap.xml, which pointed at /untrace/ — a URL that does not exist. - Regenerated the Open Graph image to match the new design. Design - Vendored the design system at src/tokens.css and composed the app from it. The old styling broke most of its rules: gradients, font weights 800 and 900, raw #2563eb, Inter, and emoji as the only status signal. - Reworked the tool into a Scan / Review / Download flow that lists what it found before changing anything, with the primary action pinned within thumb reach on small screens. - Dark by default with a light theme, applied before first paint so the page never flashes the wrong surface. - Replaced the floating donation widget, which painted an off-token circle over the page and covered the mobile action bar, with a footer link. Structure - Moved detection into src/lib/metadata.js and removal into src/lib/clean.js; the canvas callbacks are now a promise chain. - Dropped Tailwind and PostCSS. Both were configured twice, and no Tailwind utility was used. Stylesheet is 6.2 kB gzipped with the whole system in it. Fixes - The stepper never advanced: stepState(index) read `current` from the closure, which Svelte 4 cannot track as a template dependency. - The cleaner's fallback returned the original, uncleaned file while the UI reported "All EXIF metadata removed". It now reports the failure instead. Verified against a JPEG carrying EXIF, GPS, XMP and an APP11 JUMBF C2PA manifest: the output retains only APP0 and the sRGB ICC profile, with every provenance marker gone. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018pRSdR57Ze9M2meDfYZgYE
2.9 KiB
Untrace
Remove content credentials from an image.
Untrace strips C2PA content credentials, AI generation tags, GPS coordinates and EXIF metadata from a photo. It runs entirely in the browser — the file is read and rewritten in the tab, never uploaded.
What it removes
| C2PA content credentials | Manifests, claims, assertions and signatures from the Content Authenticity Initiative |
| AI generation tags | ChatGPT, DALL·E, Midjourney, Stable Diffusion, Adobe Firefly, Google Imagen |
| JUMBF containers | The APP11 and APP2 boxes that carry provenance inside a JPEG |
| GPS location | Latitude, longitude, altitude, bearing and their timestamp |
| Camera and device | Make, model, lens, serial numbers, exposure and shooting settings |
| Timestamps | When the photo was taken, digitised and last modified |
| XMP and IPTC | Creator names, captions, keywords, copyright, editing history |
| Comments and thumbnails | User comments, maker notes, embedded previews |
| PNG text chunks | tEXt, iTXt and zTXt blocks, including generation prompts |
Supported input: JPEG, PNG, TIFF, HEIC, AVIF. Output is JPEG, or PNG when the original was a small PNG.
How it works
Detection and removal are two separate passes.
Detection (src/lib/metadata.js) reads the file twice: once through
exifr for structured segments, and once
as raw bytes to find C2PA strings and JUMBF boxes that exifr does not surface.
The results are merged and bucketed into named sections, then turned into the
findings list the review step renders.
Removal (src/lib/clean.js) decodes the image to pixels on a canvas and
re-encodes it. Nothing carries over, so every metadata container is dropped at
once rather than deleting known tags one by one and leaving the rest behind.
The output is then re-read to confirm it is clean; if anything survived, a
second pass forces JPEG on a white matte.
If the browser cannot decode or re-encode the file, the original bytes are returned and the interface says so — it never reports a clean that did not happen.
Design
The interface follows the
RealBrain Design System v2.0. Tokens are
vendored at src/tokens.css; refresh with:
curl -o src/tokens.css https://realbrain.cc/design-system/tokens.css
AGENTS.md carries the rules, the self-check commands, and this repository's
open token proposals. Read it before touching any CSS.
Development
npm install
npm run dev # vite dev server
npm run build # production build to dist/
npm run preview # serve the build
Stack: Svelte 4, Vite 5, exifr. No CSS framework — the design system is plain CSS.
Who it is for
Journalists protecting sources, activists maintaining anonymity, photographers selling stock, and anyone who would rather a photo not carry its own history.
License
MIT.