TL;DR: I shipped pdfdark.org — a browser-side PDF dark mode converter. Files don't get uploaded; the entire conversion happens in your browser via PDF.js + a Web Worker. Open source (MIT), free, no signup. Reading long PDFs at night was killing my eyes. The two paths I had both sucked: Existing "dark mode PDF" web tools → required uploading the file. For research papers, contracts, medical records — that felt sketchy. No way to verify what they did with my data. OS-level "invert colors" → wrecked photos and charts. Faces became X-rays. Graphs became noise. So I built one with two non-negotiable defaults: Nothing leaves your browser. PDF.js parses the file, a Web Worker does the dark-mode pass, pdf-lib stitches the result. Verify it yourself in DevTools → Network. Images keep their original colors. A saturation classifier detects photos and figures and leaves them untouched while it darkens text and UI. Drop PDF → PDF.js renders pages to canvas → Web Worker classifies each pixel by saturation ├─ saturated pixels (images, charts) → preserved └─ low-saturation pixels (text, UI) → themed → pdf-lib assembles a new PDF → User downloads The classifier runs on OffscreenCanvas inside a Web Worker, so the UI thread never blocks on large PDFs. Output is a real PDF (image-based, one JPEG per page), so the dark mode persists when you email it, sync it to iPad, or open it on a Kindle. Not a viewer toggle. Domain: $7.50 (Cloudflare Registrar, .org) Hosting: $0 (Vercel free tier) Email forwarding: $0 (Cloudflare Email Routing) Error monitoring: $0 (Sentry free tier, errors only with full content masking) Time: a few weekends Total cost: $7.50. Whether the "no-upload" angle resonates beyond privacy nerds Edge cases that break the algorithm (weird embedded fonts, scanned PDFs) Whether to add a vector-preserving mode for text-only PDFs (current output is image-based, so text isn't selectable) Links Live: pdfdark.org GitHub: github.com/1436941541/pdf-dark Privacy: pdfdark.org/privacy