No uploads · No attachments kept · No user information retained · No accounts · No watermarks · No page limits · No catches
A NavigatorsLab project.
The pitch, in one sentence: PDF Studio is the PDF editor you keep trying to find and never find — one that actually rewrites the text already in the document, fills real forms, OCRs scans, signs, redacts, reorganizes pages, diffs two revisions, and answers questions about the file with an on-device AI — and it does all of it without your document ever leaving your computer, for free, as open source (MIT).
Every "free" PDF editor online hits you the same way: upload your contract or your ID to somebody's server, edit three pages, then hit a paywall, a watermark, a page cap, or a sign-up screen at the exact moment you press Save. Desktop editors that genuinely modify existing text cost money or lock the feature behind a subscription.
PDF Studio's answer: a full editor that runs 100% client-side. Your files are opened, edited, and saved by your own browser, on your own machine. There is no server to upload to. There is no account. There is no watermark, no quota, no "pro" tier. It is MIT-licensed open source — fork it, self-host it, ship it.
Every claim below is demonstrated by a real screenshot in this repository, captured from the running app.
The feature other "editors" fake. Pick the Edit tool and every line of the original text lights up as a clickable target. Click a line, retype it, done — the old glyphs are painted over with a background color sampled from the page itself (so it works on scanned or textured pages, not just white), and your replacement is stamped at the same position and size. On export it's real vector text in real PDF content, not a screenshot pasted on top.
- Works on any digital PDF with a text layer, at any zoom level
- Hover hints show exactly what's editable before you commit
- Exports as true vector text — searchable, selectable, print-crisp
Open a fillable PDF and the Forms dialog lists every field: text inputs, checkboxes, radio groups, dropdowns — with jump-to-page buttons. Type your values right in the list, or flip on "show fields on pages" and click the widgets directly on the page canvas. On save, values are written into the real form fields — or flattened into page content so nobody downstream can change them. Tax forms, applications, contracts: filled privately, on your machine.
Draw your signature on the built-in pad (exported as a transparent PNG), place it anywhere, resize it. Highlight, underline, strike through, freehand pen, sticky notes, text boxes, image stamps — plus styling shapes: arrows & pointers (arrowhead lands where you release), outlined rectangles, and ellipses for circling the important part. Every mark lives in true PDF coordinates and exports as real vector content. Color and line weight apply across pen, shapes, and arrows.
Drag thumbnails to reorder. Rotate 90° (honoring the file's own rotation metadata — with a one-click "Turn upright" banner when a file was scanned sideways). Duplicate, delete, insert blank pages. Merge another PDF at any position — or just drop a file onto the canvas. Split by page range (1-3,5) or explode to one-file-per-page.
At export time: page numbers (position, 1 / Page 1 / 1 of N formats, custom first number, skip-the-cover), diagonal watermarks with size/opacity/color, and header/footer lines with {page}, {pages}, {date}, {title} tokens. Rendered at true PDF coordinates — crisp at any zoom.
Full-text search across every page, case-insensitive, with highlighted matches on the page and Enter/Shift+Enter navigation with a live counter.
Drop two versions of a document into Compare and get an LCS line-diff of their extracted text — additions in green, removals in red, per-page counts. Did the payment terms change between v1 and v2? Now you know in seconds.
The ✦ assistant runs a real LLM (Qwen2.5 0.5B / Llama 3.2 1B) in your browser via WebAssembly. Ask questions, get summaries. The document text never leaves the device — the model comes to your data, not the other way around. Weights download once (~0.6–0.9 GB), then work offline. The lighter Extract text path needs no download at all and runs fully offline.
Scanned page with no text layer? The OCR tool runs Tesseract locally in WebAssembly — each recognized line becomes a real, editable text box whose background is sampled from the page, so the export turns a flat scan into clean vector text. One-time ~15 MB language download, cached forever after.
- Sideways scans: detects camera-style rotation metadata and offers a one-click, undoable "Turn upright" for every affected page — including files stuck at 180°.
- Mirrored content: some generators store artwork flipped (your email print-to-PDF might be upside down in every viewer). Flip any page horizontally/vertically — the export mirrors the real page content while your marks stay put.
- Offset page boxes: PDF Studio's geometry engine is CropBox-aware — it handles files whose MediaBox doesn't start at (0,0) (extremely common in print-to-PDF output) that displace clicks and annotations in lesser editors.
The Redact tool paints opaque boxes over sensitive content; in the exported file the marks are flattened black-over-content, so what's underneath stays hidden in the shipped document. Whiteout covers with the page-sampled background instead, for cleaning up scans invisibly. Drag, done — no "pending redaction" metaphysics like Acrobat: what you see covered is what everyone else gets covered.
The whole editor reflows for mobile: the tool rail becomes a bottom bar with labeled icons (thumb-reachable), the top bar scrolls horizontally, pages stack full-width, and touch dragging marks/pages just works. It's a PWA, so you can install it to your home screen and keep editing offline.
Everything — open files, page operations, marks, form values — autosaves to on-device storage. After a crash or refresh, the start screen offers Resume session, and picks up exactly where you left off.
It's a PWA: install it to your dock/desktop and it keeps working with the network off. Your toolkit shouldn't have an "out of service" state.
| Question | Answer |
|---|---|
| Where do my files go? | Nowhere. There is no upload endpoint. Open the devtools network tab and watch it stay silent while you edit. |
| Do you keep my attachments or user information? | No. No attachments are retained, no user information is collected — no account, no email, no analytics on your documents. Your session autosaves locally in your browser and never leaves it. |
| Watermarks? | None. |
| Page/task/hour limits? | None. |
| What downloads over the network? | The app itself, plus optional public model weights (AI ~0.6 GB, OCR ~15 MB) — cached by your browser after the first fetch. Your documents never transit the network. |
| Can I air-gap it? | Yes. Clone the repo, host dist/ on an internal network, point the AI/OCR loaders at internal mirrors. |
- Edit-text targets PDFs with a text layer; for flat scans you use the OCR tool (which exists precisely for that).
- Replacement text is set in Helvetica metrics — extremely close, but not glyph-identical to exotic embedded fonts.
- Compare diffs extracted text, not rendered pixels — it's built for contract/revision review, not detecting a shifted logo.
- The AI runs small models (0.5B–1B parameters) — good at summarizing and locating content in a document, not a replacement for a frontier model.
git clone https://github.com/Kayforkind/NavigatorsLab-PDF-Studio
cd NavigatorsLab-PDF-Studio
npm install
npm run dev # http://localhost:5199 → click "Open the demo document"
npm test # 24 unit + integration tests
npm run typecheck
npm run build # production build + PWA service worker → dist/Stack: TypeScript · React 19 · Vite · pdf.js (render) · pdf-lib (write) · Tesseract.js (OCR) · WebLLM (AI) · vite-plugin-pwa. Zero backend.
Architecture in one line: every mark lives in the page's PDF user space; the on-screen overlay and the exporter share the same CropBox-aware transform math (src/lib/viewport.ts), pinned in tests against pdf.js's own PageViewport — including offset MediaBox/CropBox origins that break most home-grown editors.
Test suite (24): document model ops & undo semantics · viewport parity with pdf.js at 0/90/180/270° · real export round-trips re-parsed with pdf.js (rotation, blanks, flattened annotations, form flattening, stamps) · LCS diff engine.
npm run build # → dist/The build uses a relative base, so the same dist/ works at a domain root, a subpath (/repo-name/), or a CDN — verified under a simulated GitHub Pages subpath. A GitHub Actions workflow (.github/workflows/deploy-pages.yml) publishes to Pages on every push to main; wrangler.jsonc deploys to Cloudflare Workers (navigatorslab.com) with npx wrangler deploy.
MIT — free for personal, commercial, and everything-in-between use. That's the whole point: LICENSE.
Free. Open source. Private by architecture, not by policy.