_ _ _ ____ ___ ____ ____ _  _ ____ ____ _  _ ____    ____ ____ _  _ ____ _  _ ____ ____
| | | |__|  |  |___ |__/ |\/| |__| |__/ |_/  [__  __ |__/ |___ |\/| |  | |  | |___ |__/
|_|_| |  |  |  |___ |  \ |  | |  | |  \ | \_ ___]    |  \ |___ |  | |__|  \/  |___ |  \

watermarks-remover

Agent skill + stdlib Python service that strips multi-vendor AI provenance marks from text and files. For privacy and hygiene on content you own.

The skill is a thin HTTP client — the agent host needs no Python. All work runs in the service.

Author: ShadowAqueduct

Latest release: v0.5.0

Covers class-level marks from Claude, Gemini/SynthID-Text, OpenAI provenance surfaces, and open-LLM schemes (Kirchenbauer green-list, keyed-Gumbel / Aaronson EXP).

Install

Skill ships no code — it calls the service over HTTP. Install the skill, start the service, set WATERMARKS_SERVICE_URL if it is not http://127.0.0.1:8765.

One installer (Python 3.10+, stdlib only):

python3 install_skill.py --skill remove-ai-marks --target claude-code

Shipped skills: remove-ai-marks (full, service-backed) and clean-user-facing-text (text only, self-contained). Use --list to see them. Existing installs are kept as backups unless you pass --force. --link symlinks the checkout for live edits.

Claude Code plugin (marketplace)

/plugin marketplace add guillaumemeyer/watermarks-remover
/plugin install watermarks-remover@watermarks-remover

Skills load as /watermarks-remover:remove-ai-marks and /watermarks-remover:clean-user-facing-text. Update with /plugin marketplace update watermarks-remover.

Grok

mkdir -p ~/.grok/skills
ln -sfn "$(pwd)/skills/remove-ai-marks" ~/.grok/skills/remove-ai-marks

Start the service

make serve                 # http://127.0.0.1:8765
# or:
python3 service/scripts/server.py --host 127.0.0.1 --port 8765

Optional system tools (used when present): c2patool, exiftool, qpdf. Core scripts need only Python 3.10+ stdlib.

Automatic cleaning via hook

A skill is an instruction — the model decides whether to run it. A hook runs on every matching tool call and does not need model cooperation.

The plugin registers a PostToolUse hook on Write|Edit|MultiEdit|NotebookEdit that runs service/scripts/hook_written_file.py:

Set mode via plugin settings (Hook mode) or WATERMARKS_HOOK_MODE=clean. Without the plugin, add the hook yourself in ~/.claude/settings.json.

Hooks cover files the agent writes and the pre-commit gate. Chat transcript text still depends on the skill (best-effort).

Quick use

SCRIPTS=service/scripts

python3 "$SCRIPTS/inspect_file.py" draft.md
python3 "$SCRIPTS/clean_file.py" draft.md -o draft.cleaned.md
python3 "$SCRIPTS/clean_file.py" photo.png -o photo.cleaned.png
python3 "$SCRIPTS/clean_file.py" notes.docx -o notes.cleaned.docx

# Text Layer A
python3 "$SCRIPTS/inspect_text.py" draft.md
python3 "$SCRIPTS/clean_text.py" draft.md -o draft.cleaned.md --stats

# Layer B rewrite (default: print prompt only)
python3 "$SCRIPTS/rewrite_text.py" draft.md --backend print-prompt --strength paraphrase

Text tools refuse binary input (DOCX, PDF, images) and point you at inspect_file.py / clean_file.py. Unrecognized formats are never auto-cleaned.

HTTP service

Same machinery as a stdlib HTTP server (service/scripts/server.py):

WM="http://127.0.0.1:8765"
curl -s "$WM/health"
curl -s -X POST "$WM/clean" -H 'Content-Type: application/json' \
  -d "{\"file\": \"$(base64 < notes.md | tr -d '\n')\", \"name\": \"notes.md\"}"

Set WATERMARKS_SERVER_API_KEY to require bearer auth. Binds loopback by default.

Detection is separate from cleaning. Text detectors (MarkLLM, keyed-Gumbel, Claude seam) and image SynthID scoring are opt-in and fail-soft.

Docker

make docker-core-build
docker run --rm -p 127.0.0.1:8765:8765 --read-only --tmpfs /tmp watermarks-remover

docker compose up -d                         # core only
docker compose --profile harness up -d       # + markllm / markdiffusion
docker compose --profile heavy up -d         # + ctrlregen / synthid (local builds)

Published images on GHCR: core, markllm, markdiffusion. CtrlRegen and SynthID scorer stay local-only (upstream licensing). Copy .env.example → .env for optional config; nothing is required for basic text cleaning.

File formats

PDF needs qpdf for a real strip (exiftool alone is incremental and leaves recoverable bytes). Ghostscript handles metadata inside embedded images. Soft-bound C2PA and pure pixel/audio/video watermarks remain out of scope for the core path.

Optional backends

Bootstrap scripts live under service/scripts/ (setup_synthid.sh, setup_ctrlregen.sh, setup_markllm.sh, setup_markdiffusion.sh). Layer B rewrite is iterative and can be driven by these detectors when configured.

How text marking works

• Layer A removes edit-based Unicode carriers (testable, lossless).
• Layer B attacks sampling watermarks via heavy rewrite (best-effort; costs style and voice).
• File cleaners strip C2PA / XMP / props from supported containers.

No tool can certify that a vendor detector will fail. Prefer a non-origin model for Layer B so you do not re-stamp the text.

Skip Layer B when quality matters more than hygiene: use Layer A + file cleaners and keep the original prose.

Pre-commit

# .pre-commit-config.yaml
repos:
  - repo: https://github.com/guillaumemeyer/watermarks-remover
    rev: v0.5.0
    hooks:
      - id: watermarks-remover-check   # fail on marks
      # - id: watermarks-remover-clean # opt-in: clean in place

Ethics

For privacy and research on content you own or are authorized to process. Not for academic fraud or false “human-written” claims. Users must follow local law. The authors disclaim liability for misuse.

See skills/remove-ai-marks/references/ethics.md.

License

MIT — see LICENSE.

Bibliography (selected)

• How Claude marks AI-generated content
• Dathathri et al., Scalable watermarking for identifying large language model outputs (SynthID-Text, Nature 2024)
• Kirchenbauer et al., A Watermark for Large Language Models (arXiv:2301.10226)
• C2PA / c2patool
• Liu et al., Image Watermarks are Removable Using Controllable Regeneration from Clean Noise (CtrlRegen, ICLR 2025)
• THU-BPM/MarkLLM, MarkDiffusion