Why most online PDF tools upload your file
Two tools can produce byte-identical output while doing completely different things with your document. The difference is architectural, and it decides who can see the contents.
Last updated: 20 November 2026
The default architecture
A conventional online PDF service is a web form in front of a server. You pick a file, the browser
posts it as multipart/form-data, the server writes it to a temporary directory, runs a
toolchain over it (Ghostscript, qpdf, pdftk, ImageMagick, sometimes LibreOffice), and streams the
result back. It then deletes the file — usually immediately, sometimes after a retention window.
That design is not malicious. It is simply the only way to run heavy native tools from a web page. It also has real advantages: a server can use a 200 MB toolchain that would never fit in a browser tab, and it can compress a file to a fifth of its size in ways a browser cannot.
But the trade is exact and unavoidable: for the duration of the job, your document is on a machine you do not control, and the operator of that machine could read it if they wanted to. Most will not. Some cannot, because the pipeline never persists anything. You have no way to tell which is which.
What a browser can do instead
A modern browser is a small operating system with a PDF engine in it. Two libraries do most of this work:
- PDF.js (Mozilla, Apache-2.0) parses the object structure, rasterizes pages to a canvas, and reads the text layer. This is the same engine that renders PDFs inside Firefox.
- pdf-lib (MIT) edits the document: copies pages between files, deletes and reorders them, sets rotation flags, embeds images, draws text, and serializes a new file.
Between them, every page-level operation on this site is possible without a network request. Merging
is copyPages. Deleting is a page-tree removal. Rotating writes an integer. Adding page
numbers draws text in the standard font. None of those need a CPU you do not own.
What local processing genuinely cannot do
Being straight about the limits is more useful than a slogan. Three things are out of reach:
- Aggressive compression. The big wins come from re-encoding images with downsampling and subsetting fonts — a native toolchain. In the browser you can rebuild the file structure (small win, often none) or rasterize pages (big win, but the text layer is destroyed).
- PDF to Word. Reconstructing layout requires a model that identifies headings, table cells and image placement. That model is too large to download for one page view.
- OCR. Recognition models are tens of megabytes and need a runtime; shipping one would make the first page load unusable on a phone.
So the honest framing is not "local is always better". It is: for the operations that are pure structure manipulation — which is most of what people actually need — local wins on privacy at no cost in quality. For the operations that need a model or a native toolchain, local cannot compete, and a site claiming otherwise is either uploading anyway or producing bad output.
How to verify a local-processing claim
Any site can write "your files never leave your device". There are three checks that take a minute:
- Watch the network. Open developer tools, switch to the Network tab, filter to
Fetch/XHR, and run the tool. A server-based converter shows a large POST — often hundreds of kilobytes to tens of megabytes, matching your file size. A local tool shows nothing but analytics. - Pull the plug. Load the page fully, then disconnect from the network (or switch the browser to offline mode) and run the operation. If it completes, the work is on your machine. If it hangs or errors, something was being sent.
- Search the bundle. Look at the JavaScript the page loads. A local tool ships pdf-lib or PDF.js, which you can find by name. A thin client just posts to an endpoint.
Why this matters more for some documents
For a conference slide deck, an upload is a shrug. For a signed contract, a medical letter, a scan of an ID document, an unpublished financial statement or a job applicant's CV, it is a different calculation — and often one covered by an obligation, not just a preference. GDPR Article 32 requires appropriate technical measures for personal data; a supplier that receives and stores the document is a processor you have to account for. Keeping the file in the browser removes that question entirely.
That is the whole argument. Not that servers are untrustworthy, but that the cheapest way to protect a document is to never send it.
Where to go next
- Merge PDF files — the clearest example of a pure structure operation.
- Compress a PDF — including an honest account of what each mode can and cannot achieve.
- What a PDF reveals about you — the metadata most people never look at.