ConvertBox Logo
ConvertBox

How browser-based file tools work without uploading files

A browser tool can read a file you pick, process it in the tab's memory and hand back the result without any upload. The price is that your own device supplies the memory and the processing power, which is why large video is the hard case.

ConvertBoxPublished

Reading a file without sending it anywhere

When you pick a file in a web page or drop one onto it, the browser hands the page a File object: a handle to that one file, not to your disk. Through the File API the page's script can read its bytes into memory. Nothing in this step involves the network; uploading is a separate, explicit act in which the script puts those bytes into a request to a server. A client-side tool simply never makes that request. It processes the bytes in memory, wraps the result in a Blob, and offers it through a temporary object URL that the browser saves like any other download.

This can be checked instead of believed. Open the browser's developer tools, switch to the Network tab and run a conversion: a tool that works locally shows no request carrying your file.

The engines: WebAssembly, Canvas and JavaScript libraries

The work itself is done by three kinds of machinery. WebAssembly is a compact binary format that browsers execute at close to native speed inside the same sandbox as JavaScript; it lets long-established C and C++ programs be compiled for the web. Canvas is the browser's own drawing surface, which can decode an image, draw it at a new size and encode the pixels again. Plain JavaScript libraries cover formats whose structure can be edited without heavy computation.

One side effect is worth knowing: a canvas holds only pixels, so an image re-encoded this way loses its metadata. ConvertBox's image compressor, for instance, removes metadata including GPS position, and scales any image whose longest side exceeds 1920 pixels down to that size.

JobTechnologyWhat happens
Video and audio conversionFFmpeg compiled to WebAssemblyStreams are decoded and re-encoded in software inside the tab
PDF encryption and decryptionqpdf compiled to WebAssemblyThe document's content is encrypted or decrypted with your password
Image compression and resizingCanvasThe image is decoded, scaled and re-encoded as JPEG, PNG or WebP
Merging and splitting PDFsA JavaScript PDF libraryPage objects are copied into a new document without re-rendering
Background removal, OCRNeural network models run through WebAssemblyA model evaluates the pixels on your own processor

Memory: the real limit of a browser tab

A native program can stream a large file from disk a piece at a time. Browser tools usually hold the whole input in memory, then the output beside it, and often an intermediate form as well. Compressed files expand when decoded: a 12-megapixel photo of 3 MB is 4000 x 3000 pixels at 4 bytes each, 48 MB once drawn on a canvas. A tab that asks for more than the browser will grant fails, and on phones the system may simply reload the page.

Upload limits in such tools are therefore memory budgets, not arbitrary rules. On ConvertBox, PDF merge takes files up to 100 MB each and keeps every source in memory with the growing output; the video converter rejects files over 50 MB; the video compressor accepts up to 500 MB but notes that files that large may fail on phones and low-memory laptops. A job that fails on a phone can pass on a desktop with more RAM.

Speed depends on your device, and video is the hard case

Since your processor does the work, the same page is fast on a recent laptop and slow on an old phone. For most jobs this is hardly noticeable: copying PDF pages or re-encoding one photo takes moments. Recognition and segmentation models are heavier, and video encoding is heavier by orders of magnitude. One minute of 1080p video at 30 frames per second is 1,800 frames of about 2 million pixels each, and every one has to be decoded, analysed for motion and compressed again.

Native encoders lean on hand-tuned processor instructions and on the hardware encoders built into graphics chips. FFmpeg in WebAssembly runs in software inside the sandbox and has neither at full strength, so it is far slower than a native build. ConvertBox's video compressor re-encodes with x264 at the medium preset, which compresses efficiently but slowly; short clips are realistic, a full-length recording is not.

First-use downloads: engines and models

Local processing does not mean the page arrives complete. Large engines are fetched only when a tool needs them, so the first run can involve a sizeable download in the opposite direction: code and models coming to you, not your file going out. The FFmpeg core is roughly 30 MB. ConvertBox's background-removal tool fetches its segmentation model and runtime from the library's public content delivery network the first time you press the button, so that run needs a connection and takes longer. OCR needs a separately trained model for each language.

Afterwards the browser cache normally serves these files, which is why a second run starts faster. On a metered mobile connection it is worth knowing that the first video or background-removal job costs data even though the file itself is never transmitted.

What works offline means, and the two features that are not local

Once a tool's page and engine are loaded, the processing needs no connection; that much of the offline claim is true. It does not make a website an installed application. Opening or reloading a page needs the network, an engine that is not yet cached cannot be fetched without it, and a browser tool should not be relied on in a place with no signal unless that exact tool was loaded beforehand.

Nor is every feature local, and a blanket statement that nothing leaves the device would be wrong. On ConvertBox two features send data out. Downloading audio from Text to Speech transmits the text, through the site's server, to an online speech service that returns the audio; playback with a voice installed on your device does not. Audio transcription relies on the browser's speech recognition, which some browsers run online: Chrome sends the captured audio to Google's speech service. Confidential text or recordings should not go through those two paths. The other tools process files in the browser as described above.

When a desktop app or a server-side service is the better choice

Browser processing wins when the file is sensitive, the job is small to medium, and installing software is not worth the effort. It loses in identifiable cases.

  • Long or high-resolution video, HEVC or AV1 output, or hardware-accelerated encoding: use a desktop encoder such as HandBrake or a native FFmpeg install.
  • Files larger than a tab can hold, or hundreds of files in one batch: use desktop software that streams from disk and can run unattended.
  • Exact control, such as print-resolution PDF rendering, downsampling the images inside a PDF, or manual masking for hair-critical cut-outs: dedicated editors expose settings that a one-screen tool does not.
  • A weak device and a file that is not confidential: a server-side service does the computing elsewhere, at the cost of upload time and of the file residing on someone else's machine.
  • Work shared between people, such as review links or team storage: that needs a server by definition.

Tools mentioned in this guide

Related guides

All guides