Free · Open Source · Runs 100% Locally

Claude's watermark is invisible. So is where your text goes when you clean it: nowhere.

Since August 2026, Claude quietly embeds machine-readable marks into everything it writes. MarkClean finds the hidden characters, shows you exactly what it found, and removes them. Everything happens on your own device.

Run it yourself in one line, even if you have never used Git: paste Clone github.com/Rjgyana/MarkClean-ClaudeWatermarkRemover and start it for me into Claude Code.

Live example: a sentence copied from Claude

Our team shipped the onboarding redesign this week.

Looks like ordinary text. Try revealing it.

Detect, then clean

Shows you the exact hidden characters found. Not a blind guess.

Nothing uploaded

Runs on your own machine. No server ever sees your text.

Free and inspectable

Open source under a personal-use license. Read every line.

Works beyond Claude

Catches the same invisible-character techniques used by ChatGPT and others.

Why This Matters Now

Anthropic made watermarking mandatory, with no opt-out

In August 2026, Anthropic began embedding machine-readable marks into Claude's output across every plan, with no setting to turn it off. Part of that is invisible Unicode woven directly into the text: zero-width spaces, word joiners, and similar characters that render as nothing but persist through copy and paste.

The announcement landed badly. Writers using Claude to proofread their own work, developers worried about marks showing up in shipped code, and anyone pasting Claude output into a CMS or a client deliverable found themselves carrying an invisible signature they never agreed to.

MarkClean exists to give you back a plain, unmarked copy of your own text, verified locally instead of taken on faith.

0

Requests sent to any server when you run Layer A detection and cleaning. The entire scan happens on your device.

How It Works

Two layers, and you choose how far to go

Layer A runs instantly with zero setup. Layer B is optional, local, and entirely under your control.

Layer A / Step 1

Detect

Scans your text against a table of known invisible-character watermarking techniques: zero-width spaces, joiners, bidirectional-control characters, tag characters, and more. Shows you exactly what it finds and where.

Instant · no setup

Layer A / Step 2

Clean

Strips the confirmed watermark characters while carefully preserving emoji, line breaks, and everything else about your text exactly as written.

Instant · no setup

Layer B / Optional

Rewrite locally

For statistical, word-choice-level watermarking that has no invisible character to find, you can optionally rephrase the cleaned text through a free local AI model via Ollama. Nothing leaves your device.

Optional · you decide

MarkClean vs. Typical Online Removers

Most “watermark removers” never actually check

Their own documentation says it plainly: without a public detector, they rewrite blind and hope. MarkClean checks first.

MarkCleanTypical online remover
Shows you which characters it foundYesNo, rewrites blind
Your text leaves your deviceNeverSent to their server
Source code you can readFully openClosed
CostFree, alwaysFree tier, often paid
Rewrite modelYour own local modelTheir server-side model

Good To Know

Frequently asked questions

Starting August 2026, Anthropic embeds machine-readable marks into Claude's output. Some are literal invisible Unicode characters (zero-width spaces, word joiners, and similar) woven into the text itself. They are undetectable to the eye but readable by software.

If this is useful, a star helps more people find it

MarkClean stays free and open because people like you point others to it. One click costs nothing and keeps this maintained.

git clone https://github.com/Rjgyana/MarkClean-ClaudeWatermarkRemover.git