Open source
Claude watermark remover on GitHub: five free paths, and what each one returned
Search for a Claude watermark remover and the first thing you meet is a free repository, a skill and a prompt rather than a service. Every one of them is a reasonable place to start. We ran them on the same marked texts; below is what came back and when the free path is the right answer.
Runs made on 12 September 2026 on three texts carrying a watermark of the published class, placed with a key of ours and read by a detector that holds it. Versions, prices and word caps move fast in this niche; the date tells you how much to trust the figures.
Five free paths, and what each one returned on 12 September 2026
-
1. watermarks-remover, cleaning layer
free · MIT · runs offline on your machine
github.com/guillaumemeyer/watermarks-remover
The best known name in the niche, version 0.7.0 when we cloned it on 12 September 2026. This layer strips invisible Unicode, unusual spacing and provenance records (C2PA, EXIF, XMP) across roughly twenty formats, from images and PDF to DOCX, EPUB, spreadsheets, video and audio. That job it finishes completely, in a second.
On our samples the prose came back identical and our mark was still found in 3 runs of 3. Every figure from that run.
-
2. watermarks-remover, rewriting layer
free · you supply the model and the key
The layer that reaches word choice. It rewrites your text through whatever model you point it at, on your machine or behind someone's API. Which model does the rewriting is your decision, and a commercial assistant puts that provider's own signal back into the wording.
Our mark was not found in 2 runs of 3. The material paid for it: on the sample with the most checkable facts the run returned 28 of 32 facts, and across the samples the text came back 23 to 36 percent shorter than it went in. Every figure from that run.
-
3. Humanize Claude text: a prompt or skill inside Claude
free · runs in the Claude you already pay for
The path people find first: a skill or a system prompt that asks Claude to rewrite its own draft until it stops sounding like a model. It does what it says about tone. On the mark it hands the job back to the model that placed it: a light edit leaves most of the wording where it was, and a full rewrite is fresh Claude output, generated under the same policy as the draft.
-
4. Free web rewriters and humanizers
free · nothing to install · word caps
Paste a text, press a button, get a paraphrase back. The 4 on the open list made 8 runs between them: our mark was not found after 3 of those and found again after 4. None of them says which model rewrote your text or what happened to your figures and names. The whole table, one row per tool.
-
5. Character cleaners in the browser
free · no account on most
Strip invisible characters, long dashes and in some cases file metadata, instantly. On our samples the 2 on the open list handed the same words back in all 6 runs, and our mark was found again in every one. If hidden characters are the whole problem, it is solved here, and it is free.
Three kinds of free tool, side by side
Every free tool in the niche is one of three kinds, and the kind decides what happens to your text. Character cleaners take out what can be found and leave the wording alone. Free rewriters hand back a paraphrase from a model they do not name. The open-source remover does both jobs on your machine, with a model you bring.
| Character cleaners | Free rewriters | Open-source remover | |
|---|---|---|---|
| Price | free | free, with caps | free, MIT |
| Invisible characters | ✓ removed | usually rewritten away | ✓ removed |
| File provenance metadata | some tools | ✗ text only | ✓ ~20 formats |
| The watermark itself, which lives in word choice | ✗ untouched: 180 of 180 marked texts still detected after a free cleaner, 26 August 2026 | rewritten, never measured by the tool; our mark not found after 3 of 8 runs on our bench | rewriting layer with your model: our mark not found in 2 of 3 runs |
| Proof your meaning survived | ✗ | ✗ | ✗ no report; 36 of 43 facts came through on our run |
| Length limits | none | often 600 to 3,000 words per request | none |
| Where your text goes | their site; some advertise "not stored" | an undisclosed model provider | stays on your machine |
Every figure in the table is a run on our marked sample, and the checker behind it is free as well: code, samples and run files at github.com/Yurakonoplya/unmark-checker. Take a sample, run it through any tool you were about to try, paste the result back. Test any remover on our marked sample.
When the free path is the right answer
-
Metadata and hidden characters are the whole problem
Then it is solved: free, offline, in a second, and you can confirm it yourself.
-
Short pieces of code
Source code leaves the mark little room: few places where the wording could have gone another way. Where it sits in a repository.
-
You can host a model and judge the result
Then nothing beats it on privacy: the material never leaves your machine, and an MIT licence lets you read every line before you run it.
Prices in this niche, dated
Taken from the tools' own pricing pages on 20 and 21 August 2026. They will drift; the date tells you how much to trust them.
| Tool | Price | Model |
|---|---|---|
| Undetectable AI | $9.99 to $42.50 a month | subscription, 10,000 to 50,000 words a month |
| WriteHuman | $12 to $36 a month | subscription, 600 to 3,000 words per request |
| StealthGPT API | $0.20 per 1,000 words | pay per volume |
| HumanizerTech | from $4 | one-off credits |
| Free tiers, browser cleaners, watermarks-remover | $0 | caps per request or a model of your own |
Questions people ask
Frequently asked
Is there a free Claude watermark remover on GitHub?
Yes, and one name leads the niche: guillaumemeyer/watermarks-remover, MIT licence, version 0.7.0 when we cloned it on 5 September 2026. It is two tools in one package: a layer that strips invisible characters and provenance metadata out of files across roughly twenty formats, and a layer that rewrites the prose through a model you supply. Only the second reaches word choice, which is where the statistical mark lives.
Does the GitHub watermark remover skill actually work?
The two layers answer differently. The cleaning layer returned our marked texts with the prose identical and our mark still found in 3 runs of 3: cleaning characters is a real job, but a different object from the mark. The rewriting layer did move it: our mark was not found in 2 runs of 3, and the material paid for it. On the sample with the most checkable facts it returned 28 of 32 facts, and the text came back 23 to 36 percent shorter.