Fixlark

About

What runs where

The short version: fourteen of these eighteen tools never send your file anywhere, four of them do and say so, and the whole thing runs on one small server.

Two kinds of tool, and one way to tell them apart

Every tool on this site carries a stamp in the same place, in one of two colours.

In your browserYour file is opened by this page and never sent anywhere. It works with the network off.

Merging a PDF, resizing a photograph, counting words, diffing text, formatting JSON, making a QR code — none of that needs a server. Your browser can do all of it, faster than the round trip would take. So these tools have no upload, no file size limit set by our disk, and no cost to us in running them. You can check the claim the easy way: turn off your network connection and they still work.

On our serverYour file is uploaded, processed, and deleted — 30 minutes after the job, or 10 minutes after you download it.

Four things a browser genuinely cannot do: check grammar against a rule engine, transcribe speech, synthesise speech, and run a segmentation model. Those upload your file, do the work on a processor of ours, hand back a result, and delete what they received.

What the server actually is

One small virtual machine: four shared processor cores, 7.7 GB of memory, no graphics card, and no swap file. It also runs a dozen other things, some of which people rely on.

That shapes everything about the four server tools. There is a queue, and it runs one heavy job at a time, because two would not be faster — they would be slower, and would take processor time from everything else on the machine. When you are queued you are told your position and an estimate measured from what jobs have actually taken today. When the queue is full you are refused straight away with a number attached, because being told "come back in five minutes" is better than joining a line that will not move.

No graphics card also settles what is not here. No image generation, no video, no chatbot. Those need hardware this machine does not have, and a free site that promised them would be lying.

What is deleted, and when

  • Your uploaded file is deleted the moment the result exists. Not on a timer — immediately. A transcription that has produced its text has no further use for the audio.
  • The result is deleted 30 minutes after the job finishes, or 10 minutes after you download it, whichever comes first. Every result has a button to delete it sooner.
  • Restarting the service deletes everything on disk. A restart loses the job list, so anything left is by definition something nobody can ask for.
  • The grammar checker never writes anything to disk at all. There is no file to delete because there was never a file.
  • Nothing is backed up, and nothing is used to train anything.

Written out in full, with the rules as they are implemented and tested: what is kept, and for how long.

Languages: what is missing, and why

This site is English-first, and two of its tools cannot serve other languages at all. That is worth stating plainly rather than discovering after an upload.

Grammar checking is English only. Two separate reasons. LanguageTool has no Vietnamese rule set — none exists, and no server would change that. The other European languages do exist, and we measured what running them costs: English alone needs about 509 MB of memory; all nine together need about 1.6 GB. On a machine with 7.7 GB, no swap and other people's applications on it, that is not a gigabyte we are willing to spend.

The speech tool has no Vietnamese voice. Kokoro ships 54 voices across nine languages and Vietnamese is not among them. Feeding Vietnamese text to an English voice produces an English speaker reading Vietnamese letters, which is unintelligible.

Transcription is English only, and this one is a choice. Whisper is genuinely multilingual and handles Vietnamese well; we run the English-only model because it is more accurate on English at the same size. This is the one language gap on the site that could be closed by a decision rather than by hardware.

The tools that do not touch language at all — every PDF and image tool, the counter, the diff, the formatters, the QR generator — work identically in any language, because they never look at the words.

The full write-up, including the measurements behind those numbers and what the alternatives would cost, is published here: language coverage.

What this site is not

  • Not a place to sign in. There are no accounts, no email addresses collected, and no password fields anywhere.
  • Not a backup service. Everything is deleted. If you need a file, download it.
  • Not a replacement for professional software for colour-managed photography, print production, or anything where a document's structure matters more than its appearance.
  • Not free because you are the product. It is free because most of it costs nothing to run: the browser does the work. The four server tools have limits and a queue precisely so that they stay affordable without a subscription.

Third-party software this is built on

  • pdf-lib (MIT) — PDF structure editing, in your browser.
  • pdf.js (Apache-2.0) — PDF rendering, in your browser.
  • qrcode-generator (MIT) — QR encoding, in your browser.
  • LanguageTool (LGPL-2.1) — grammar and spelling. We run our own copy; nothing reaches LanguageTool's hosted service.
  • whisper.cpp (MIT) with OpenAI's Whisper base.en weights (MIT) — transcription.
  • Kokoro-82M (Apache-2.0) — speech. hf.co/hexgrad/Kokoro-82M.
  • U^2-Net / u2netp (Apache-2.0, Qin et al. 2020) — background removal.

All of the browser-side libraries are served from this site. Nothing is fetched from a content delivery network, because a site whose promise is "your file stays here" should not begin by calling three other companies.