Try it on a real paragraph
This live demo runs the same rules the kit ships with. Paste your own draft, or load the sample.
Step 1 Paste your draft
Rule-based. Runs in your browser. Your text is never sent anywhere and never stored.
Paste a paragraph first, or load the sample, then scan. Nothing to check on an empty box.
Step 2 · Find it The tells, marked in context
Hover or tab onto any marked word to read why it reads as AI and how to fix it. Every point of the score traces to a rule you can read.
Step 3 · Fix it & prove it The rewrite, changes in green
Mend only touches the mechanical tells, and shows every edit. Your voice and meaning stay yours.
Grey = before (54). Green = after Mend (90). The notch is the 80 "sounds human" target.
Plus the hedge "It's important to note that" was removed from the start. That is 8 safe edits in all.
Left for you (3). Mend won't guess your meaning.
- "is designed to", a marketing boast. Fix: say what it does, plainly, in the present tense.
- "not only fast but also", a correlative construction. Fix: split into two plain clauses, or just say the second half.
- "fast, clean, and reliable", a rule-of-three pile-up. Fix: keep the strongest item; vary your list lengths.
Why this one is different
It shows its work
Not a guess, not a black box. Every point of the score and every edit traces to a rule you can read. You can audit the whole thing in one file.
It restores, not just accuses
Most detectors hand you a verdict and walk away. This one comes with a fix for every finding, and tells you what it left for your judgment.
It is fully offline
Pure Python standard library. No network, no API, no third-party packages. Nothing you write ever leaves your machine.
Accessibility is built in
The report is dyslexia-friendly by design: legible type, real contrast, text-labelled findings. It scores your text for readability too.
What you get
tells.py, the detector and coachmend.py, the transparent rewriter- The full rule reference and a quick-start guide
- Worked samples and an example accessible HTML report
- The test suite (all tests green) so you can verify it yourself
Requirements: Python 3.8 or newer. That is all. Runs on Windows, Mac, and Linux.
Get the Human Voice Kit
$9 minimum · pay what you want, $15 suggested
Two small tools that work together, fully offline. Find the tells, fix the safe ones in front of you, prove the lift. Nothing leaves your machine. That is the whole point.
Checkout link goes live on launch.
Honest note
This is version 1. It catches the common, high-signal tells and fixes the safe ones. It does not pretend to read minds or to be a perfect detector, because no honest tool can. A personal voiceprint mode and a one-click web version are on the way. If it ever fails you, we fix it fast.