markscrub

CLI + universal agent skill to scrub AI provenance marks from text and files — for privacy and hygiene on content you own.

Layer	Target	How
A	Invisible Unicode, exotic spaces, bidi, tag chars	Deterministic (clean)
B	Statistical (token-sampling) text marks	Optional rewrite backends
Files	C2PA / EXIF / XMP / doc props	PNG, JPEG, SVG, PDF, DOCX, HTML, MD

Honesty: Layer A and metadata cleans are verifiable. Layer B is best-effort. No tool can certify that a vendor detector will fail.

Want a website instead of a CLI? Use Claude Watermark Remover — same project, hosted. Free in-browser check/clean (Layer A). Paid rewrite for the statistical text mark (Layer B). No install. Use this repo if you want local files, C2PA/metadata, CI, or an agent skill.

Try Layer A locally in-browser: open playground/index.html

Install

git clone git@github.com:anshaneja5/markscrub.git
cd markscrub
npm install
npm run build
node dist/cli.js help

Dev without build:

npm run markscrub -- help

Agent skill (Cursor / Claude Code / Codex)

./scripts/install-skill.sh cursor    # ~/.cursor/skills/remove-ai-marks
./scripts/install-skill.sh claude    # ~/.claude/skills/remove-ai-marks
./scripts/install-skill.sh codex     # ~/.codex/skills/remove-ai-marks

Or copy/symlink skills/remove-ai-marks/ into your agent’s skills directory, then ask to scrub AI provenance marks / /remove-ai-marks.

CLI

# Inspect (file, stdin, clipboard, directory)
markscrub inspect draft.md
markscrub inspect draft.md --json --fail-on-findings
markscrub inspect . --recursive --sarif out.sarif
printf 'Hi\u200b' | markscrub inspect -
markscrub inspect --clipboard

# Clean
markscrub clean draft.md -o draft.cleaned.md --diff
markscrub clean ./content -o ./cleaned --recursive --json
printf 'Hi\u200b' | markscrub clean - -o cleaned.txt
markscrub clean --clipboard --diff          # read + write clipboard
markscrub clean draft.md -o x.md --dry-run --diff

# Rewrite (Layer B)
markscrub rewrite draft.md --backend print-prompt
markscrub rewrite draft.md -o out.md --backend ollama --model llama3.2
markscrub rewrite draft.md -o out.md --backend lmstudio
markscrub rewrite draft.md -o out.md --backend openai --strength strong

Rewrite backends

Backend	Default endpoint	Notes
print-prompt	—	Default; no API key
ollama	http://127.0.0.1:11434/v1	Local
lmstudio	http://127.0.0.1:1234/v1	Local
openai	https://api.openai.com/v1	Needs API key

export MARKSCRUB_REWRITE_BACKEND=ollama
export MARKSCRUB_REWRITE_MODEL=llama3.2
# openai:
# export MARKSCRUB_REWRITE_API_KEY=sk-...

Prefer a rewrite model different from the suspected origin model.

CI / GitHub Action

- uses: anshaneja5/markscrub@main
  with:
    path: .
    fail-on-findings: true
    sarif-file: markscrub.sarif

Or locally:

markscrub inspect ./src --fail-on-findings --sarif markscrub.sarif

Report shape

{
  "tool": "markscrub",
  "version": "0.2.0",
  "action": "clean",
  "findings": [{ "kind": "unicode", "verifiable": true, "message": "…" }],
  "actions": ["unicode: removed 3 character(s)"],
  "diff": { "summary": ["…"], "before": "…", "after": "…" },
  "residualRisk": { "level": "low", "notes": ["…"] }
}

Coverage

Channel	Support
Unicode / edit-based text	Yes (Layer A)
Statistical sampling text	Best-effort rewrite (Layer B)
C2PA / file metadata	Yes for listed formats
PDF without exiftool	Best-effort XMP/Info scrub
PDF with exiftool	Preferred full metadata strip
Pixel image marks	Out of scope

Disclaimer

Rewriting changes voice and may reduce quality. If quality matters more than hygiene, use Layer A + file cleans only.

License

MIT