The developer subset of this site, grouped so you do not have to hunt: converters, inspectors, and the file tools that come up in a working day.
Every tool in this collection runs locally: no file and no string is uploaded. Formatting JSON, decoding Base64, inspecting a JWT, shrinking an image, merging PDFs — all of it happens in the tab.
That is not a preference, it is about what developers actually paste. The JWT you want to decode usually carries a real user identifier and a signature; the JSON response you want to pretty-print is production data. Pasting either into an online formatter turns it into a log entry on somebody else's server.
Many companies forbid external dev tools, and what the rule targets is data leaving the machine, not the tool itself. With no request, that condition is met by construction — open the network tab while you work and you can watch nothing go out.
Start with the unreadable string in the log. If it begins with eyJ it is a JWT; if it ends in = it is probably Base64. A JWT is signed, not encrypted, so the payload reads out directly — the first thing to check is whether exp and iat line up with the server clock.
Timestamps come next. A date near 1970 means seconds were read as milliseconds or the other way round. When you cannot tell whether a log line is UTC or local, converting it once settles the argument.
For configuration drift, use the diff. Put the pre-deploy and post-deploy config side by side and «nothing changed» usually stops being true right there.
With cron expressions the two fields people get wrong are the weekday and «every N minutes». */5 and 5 are entirely different schedules, and that one character is the difference between a job running 288 times a day and once.
«How many users can this server take» is not answered by a spec sheet. What you need is Little's Law: concurrency = arrival rate × average response time. That is the relationship the server load simulator works through.
The part most estimates miss is where the bottleneck sits. Doubling CPU cores does nothing if the connection pool holds 20 — the narrowest stage sets total throughput, no matter how large the stages in front of it are.
Response time does not grow linearly either. Past roughly 70% utilisation, queueing delay climbs steeply. That is why keeping steady-state utilisation near half is not waste; it is what you pay for predictable latency.
Treat the output as an order-of-magnitude check, not a forecast. It tells you whether the answer is two servers or twenty before you walk into the planning meeting. Real numbers still come from a load test.
An 8 MB screenshot rejected by the issue tracker, or three PDFs that have to be submitted as one file, are ordinary developer problems. Neither is worth installing ImageMagick for, and neither is worth uploading to a conversion site.
Compression has two levers: resolution and re-encoding quality. Dropping a 4000 px screenshot to 1600 px saves far more bytes per unit of visible quality than turning the quality slider down alone. For documentation, 1600 px on the long edge is usually plenty.
Merging PDFs and turning photos into a PDF finish in the same place. Contracts, invoices and signed forms — exactly the documents you should not upload — are the main input for those two operations, which is where the no-upload property earns the most.
No upload still leaves a screen. Decoding a production token while sharing your screen exposes it just as thoroughly as sending it somewhere.
Pasting the decoded result into a ticket or a chat is the other common accident. JWT payloads routinely contain user ids, e-mail addresses and role claims.
When testing a regular expression, use shaped sample data rather than real records. Matching a pattern needs the shape, not the values.
These are inspection tools. Decoding a JWT is not signature verification — that has to happen on the server with the key. A readable payload says nothing about whether the token is valid.
The load simulator does not replace a load test. Garbage-collection behaviour, kernel settings, network latency and cache hit rates all move the real numbers.
Lossy image compression cannot be undone. Keep the original and ship the compressed copy.