The browser can now do the work: tools that run on your device

Browser tools changed a trade that used to be fixed: convenience in exchange for uploading your file. That trade is now optional for a growing set of tasks, because the browser can do the work itself.

What changed

WebAssembly lets compiled code — the same libraries desktop applications use — run in a tab at close to native speed. WebGPU gives that code access to the graphics hardware, which is what makes model inference practical rather than theoretical.

The practical result is that these now run locally in a browser:

• Image conversion and compression, including HEIC, WebP and AVIF.
• PDF operations — merge, split, rotate, and rendering — through libraries such as pdf.js.
• Video transcoding, via ffmpeg compiled to WebAssembly.
• OCR, extracting text from images.
• Background removal and other segmentation, using small models.
• Speech transcription, with models small enough to download once and reuse.

The test, which has not changed

Load the page. Disconnect the network. Use the tool.

If it works, the processing is local and your file never left. If it fails, it was being uploaded. This takes ten seconds, requires no technical knowledge, and is more reliable than any privacy policy.

A refinement for the curious: open developer tools, watch the Network tab, and look for a request carrying roughly your file's size.

What local processing costs

It is not free, and the trade-offs are worth knowing before you are surprised by them:

• A large first load. A model or a codec has to be downloaded once — sometimes tens or hundreds of megabytes. It is cached afterwards, but the first use is slow.
• Memory. A tab has less than a server. Very large videos and very long documents can fail in ways that look like a bug and are a limit.
• Speed varies by machine. The same tool is fast on a recent laptop and slow on a five-year-old phone, because it is your hardware doing the work.
• Quality is sometimes lower. A model small enough to download is smaller than one running on a server. For transcription and background removal the gap is real, though narrowing.

Where local is clearly right

• Anything with personal or commercial information — statements, contracts, identity documents, client material, medical letters.
• Anything covered by an employer's data policy, which in many organisations is most work documents.
• Repeated small jobs, where upload and download time exceeds the processing.
• Working offline, on a plane or a poor connection.

Where a server still wins

• Very large files that exceed what a tab can hold.
• The best available quality in transcription, translation or upscaling, where model size genuinely matters.
• Batch work across hundreds of files, where a queue on a server beats a tab you must keep open.
• Anything needing data the tool does not have — a live lookup, a database, a comparison against something remote.

The question is no longer "is this site trustworthy?" for most everyday tasks. It is "does this tool need the network at all?" — and you can answer that yourself in ten seconds.

A practical habit

Keep one local tool bookmarked for each of the four things people convert most: images, PDFs, video and text. Verify each one with the airplane-mode test once, when you are not in a hurry.

Then the decision at the moment you need it is already made, which is the only time this kind of care actually survives contact with a deadline.

What are browser tools now capable of?

Browser tools are web pages that run compiled code inside the tab, which means the same libraries desktop applications use can process your file without it leaving the machine. WebAssembly made that practical and WebGPU made model inference practical on top of it.

What runs locally today?

Six categories that all required an upload three years ago now do not, which changes the privacy question for the whole class of free tools.

What does local processing cost?

• A large first load — a codec or model downloads once, sometimes hundreds of megabytes.
• Memory limits. A tab has less than a server, so very large files fail in ways that look like bugs.
• Speed varies by machine, because it is your hardware doing the work.
• Quality is sometimes lower, particularly for transcription, where model size genuinely matters.

The test is unchanged and takes 10 seconds: load the page, disconnect the network, use the tool. WebAssembly is what makes the answer "it still works" possible at all. See online tools, file tools and image tools.

Where a server still wins

• Very large files, beyond what a tab can hold in memory.
• The best available quality in transcription or upscaling, where model size matters.
• Batch work across hundreds of files, where a queue beats a tab you must keep open.
• Anything needing remote data the tool does not have locally.

Where browser tools clearly win

• Anything personal or commercial — statements, contracts, identity documents, client material.
• Anything under an employer's data policy, which is most work documents.
• Repeated small jobs, where upload and download exceed the processing time.
• Working offline, on a plane or a poor connection.

The question is no longer whether a site is trustworthy for most everyday tasks. It is whether the tool needs the network at all — and you can answer that in about 10 seconds, without reading anything the site says about itself.

Keep 1 local tool bookmarked for each of the 4 things people convert most, verify each once with the airplane-mode test, and the decision is made before the deadline arrives. That is the only time this kind of care survives contact with a real workload.

Where a server still wins

• Very large files, beyond what a tab holds in memory.
• The best available quality in transcription or upscaling, where model size matters.
• Batch work across hundreds of files.
• Anything needing remote data the tool does not have locally.

Where browser tools clearly win

• Anything personal or commercial — statements, contracts, identity documents, client material.
• Anything under an employer's data policy, which is most work documents.
• Repeated small jobs, where upload and download exceed the processing.
• Working offline.

The question is no longer whether a site is trustworthy for most everyday tasks. It is whether the tool needs the network at all — answerable in about 10 seconds, without reading anything the site says about itself.

Keep 1 local tool bookmarked for each of the 4 things people convert most, verify each once with the airplane-mode test, and the decision is made before the deadline arrives.