watermark-cleaner

Watermark Cleaner is a deterministic, offline, zero-dependency command-line tool for cleaning content you own before publication. It removes mechanical traces that can mark text as machine generated, flags stylistic AI phrases, and strips identifying metadata from supported image and document formats without altering pixels or legitimate text. The project ships a Python CLI and a Node CLI that share one rule set and are tested to behave identically. The tool can scan files or folders with check, which reports findings without changing anything and exits non-zero when blocking issues are present, making it suitable as a publish gate. The fix command cleans in place and writes a .bak backup next to changed files. It supports invisible Unicode cleanup, typography normalization, AI phrase detection, and lossless metadata stripping for JPEG, PNG, WebP, SVG, DOCX, and ODT. GIF and TIFF metadata are reported but not modified, and PDF support is not yet available. It is careful about what it preserves: zero width joiners and non-joiners needed for emoji or right-to-left scripts, bidi controls in RTL documents, non-breaking spaces in numbers and abbreviations, variation selectors, and genuine Cyrillic or Greek text are kept unless an aggressive mode is used. Markdown code blocks and YAML frontmatter are protected from typography and voice rules, while hidden characters inside code are still removed as a defense against Trojan Source-style attacks. The project can be installed through pipx, npm, uvx, npx, or from source, and can be integrated with pre-commit, git hooks, or deploy gates. Configuration is provided through a JSON file, allowing users to control voice rules, homoglyph handling, dash replacement, banned phrases, ignored phrases, supported text extensions, and excluded paths. An optional DeepL rewrite command is available for statistical watermark degradation, but it is off by default and sends text to DeepL, so it should not be used with confidential content. Watermark Cleaner is explicitly not a detector or certification tool. It does not remove pixel-domain image watermarks, training-time backdoors, or statistical token-level watermarks deterministically. Its purpose is hygiene, privacy, and transparency: inspect hidden content, remove unwanted machine fingerprints, and publish cleaner files.