I built 15 developer tools that never touch a server — here is what the local-only parts actually cost
I am a student in Korea, and I kept running into the same small annoyance: every time I needed to compress an image for a blog post, decode a JWT, or check a color contrast ratio, the fastest search result was a site cov
I am a student in Korea, and I kept running into the same small annoyance: every time I needed to compress an image for a blog post, decode a JWT, or check a color contrast ratio, the fastest search result was a site covered in ads that wanted me to upload the file first.
Uploading a screenshot of an internal dashboard to a random image compressor is not something I want to do, and it is not something a company should do either. So I started building the tools I actually use, with one rule: nothing goes over the network. Everything runs in the browser tab, on the user's own machine.
I have been shipping one tool a day, and the hub now has 15 of them (16 pages counting the standalone builds). This post is about the parts that were harder than they looked, rather than a feature list.
The rule, and where it breaks
The constraint is simple: a tool may use the DOM, Canvas, Web APIs, and the user's own CPU. It may not POST a file anywhere. No backend, no analytics endpoint that receives content, no "free tier" API key baked into the page.
That works surprisingly well until you hit one of three walls.
1. Anything that "just needs a library" also needs a bundle
A regex tester is easy. A good regex tester needs a highlighting engine, and pulling a 200 kB syntax highlighter into a page that otherwise ships 20 kB of inline JavaScript is a bad trade for a tool that should feel instant on a mid-range Android phone. I ended up writing the tokenizer for the SQL formatter and the diff engine by hand instead of importing one:
- SQL formatter: a small tokenizer that classifies strings, quoted identifiers, comments, and keywords, then re-emits them with indentation rules. Static syntax hints come from the same token stream, so there is no second pass.
- Diff checker: line-level LCS for the aligned view, then a word-level pass inside each changed line so a one-word edit does not paint the whole line red.
Both are a few hundred lines. That is less code than the configuration needed to tree-shake a library down to the same behavior.
2. Text encoding is not btoa
This one bit me on the Base64 tool. btoa() throws on anything outside Latin-1, which means it fails on Korean text and on emoji — the two things a Korean user will paste first.
The fix is to stop thinking in strings and start thinking in bytes:
const bytes = new TextEncoder().encode(text);
const binary = Array.from(bytes, b => String.fromCharCode(b)).join("");
const b64 = btoa(binary);
Decoding needs the mirror image (atob → Uint8Array → TextDecoder), plus tolerance for what people actually paste: data URLs with the data:image/png;base64, prefix, wrapped lines at 64 characters, stray whitespace, and the URL-safe alphabet with padding stripped. Tolerating that input is most of the work; the conversion itself is four lines.
3. Big files are a memory problem, not a CPU problem
An image compressor is trivially "draw to canvas, call toBlob('image/webp', q)". On a phone, with a 12 MP photo, the failure mode is the tab being killed while three copies of the bitmap are alive: the decoded <img>, the canvas backing store, and the encoded blob.
What helped:
- decode once, draw once, then release the source before encoding
- reset
canvas.width = 0after use so the backing store is freed rather than waiting for GC - show the encoded size against the original before offering the download, because "smaller" is the only reason anyone opens this tool
The Base64 tool has the same shape with a different ceiling: a data URL is roughly 4/3 the size of the file, and pasting a 5 MB image into a textarea produces a 6.7 MB string that some editors handle badly. So the file path uses drag-and-drop and FileReader instead of the textarea, and the page tells you the size increase before you commit.
What a JWT decoder should warn you about
Decoding a JWT is base64url plus JSON, so the interesting part is not the decoding. It is telling the user something useful:
-
alg: noneis accepted by some badly written verifiers, so flagging it matters more than pretty-printing the claims -
expin the past is common in bug reports, so a live countdown is worth more than a timestamp - Korean claims arrive as escaped Unicode, so
\uXXXXsequences need normalizing before display or the payload looks like noise
I added HS256 verification as an optional step — the signature check happens in the browser with the pasted secret, and the secret is never transmitted. Verification that requires a server would have broken the rule, so it stays optional and clearly labeled.
A note on contrast and accessibility checks
WCAG contrast is a formula, not a service: relative luminance with gamma correction, then (L1 + 0.05) / (L2 + 0.05). The part people get wrong is the threshold table — normal text needs 4.5:1, large text 3:1, and UI components 3:1 — so the tool shows all three verdicts side by side instead of a single pass/fail.
The honest limits
- No collaboration or history. Everything is per-tab. Reload and it is gone, which is the price of "we never receive your data".
- No server-side parsing. PDF and Office file input is basically out of scope; I would rather skip a tool than fake it badly.
-
Browser support.
OffscreenCanvasand WebP encoding in canvas are fine in current Chrome, Edge, Firefox, and Safari, but older Android WebViews fall back to JPEG.
Where it is going
One tool a day, chosen by what annoyed me that week. Next up is a UUID/ULID/SHA-256 generator, then whatever breaks most in the feedback.
If you have a utility you keep reaching for that only exists as an ad-covered upload page, I would genuinely like to hear it — that is how the list has been growing.
The hub is here, and every tool is free, with no signup: https://lee-know-share.blogspot.com/p/dailytools.html
No affiliate links, no tracking pixel on the tools. If a tool uploads anything, that is a bug.
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.