The tool separates verifiable cleanup from removal attempts
watermarks-remover handles several jobs that are easy to confuse. Its deterministic layer removes invisible Unicode, exotic spaces, bidirectional controls, and similar text carriers. File handlers strip C2PA, EXIF, XMP, document properties, and related metadata from supported containers. A separate rewrite layer attacks statistical text marks by changing wording, while optional research backends try known image, audio, or video schemes.
That division is the project's best judgment call. A byte removed from a file is observable. A detector result only speaks for the configured detector, model, and key. The README says MarkLLM checks are same-configuration experiments and cannot certify that a vendor's detector will fail. It also says rewriting can flatten tone and precision. If your interface reduces all of that to one green badge, you have thrown away the most useful honesty in the project.
What happened when we ran it
Our sandbox installed commit 258cf20 in 6 seconds, pulling 55 packages and occupying 74 MB. The build finished successfully in 1 second. Tests also passed, taking 96 seconds, and pip-audit reported 0 known vulnerabilities. We ran the repository in an unprivileged Debian container with Python 3.12, 3 CPUs, 8 GB of RAM, and no secrets.
The checkout contained 292 files, about 48,621 lines of source, and used 2.1 MB before installation. We found 3 CI workflow files, a compose file, and a tests directory. There was no Dockerfile at the repository root. Those results cover the commit and commands we ran. They do not measure removal accuracy, visual or audio damage, rewrite quality, or success against a private vendor detector.
The small core grows when a format needs real cleaning
The basic local path is approachable. Python starts a standard-library HTTP service on 127.0.0.1:8765, and the command-line tools can inspect or clean individual files. The published core container adds qpdf, ExifTool, and c2patool. PDF handling explains why these extras matter: ExifTool edits can leave old bytes recoverable, so qpdf performs a structural rewrite, while Ghostscript is used for metadata inside embedded images.
Layer B text work has a different bill. The /clean service needs a rewrite backend for text, such as local Ollama or an explicitly allowed remote endpoint. Optional MarkLLM, MarkDiffusion, CtrlRegen, and SynthID components add external repositories, model weights, licensing conditions, or GPU demands. Our 74 MB environment is a fair description of the tested Python setup. It is not the footprint of every research backend running together.
Hosted skills still depend on a reachable service
The Claude Code plugin can run a post-write hook in check or clean mode, and a pre-commit path can inspect files before they enter Git. The skill itself is deliberately thin. It calls the HTTP service instead of carrying the cleaning implementation. That is sensible on a developer machine where the service is on loopback and the hook can reach it.
Hosted agents expose the boundary. Open issue 356 reports that a claude.ai sandbox saw connection refused at 127.0.0.1:8765, could not start Docker there, and did not receive the referenced service files in the plugin package. The README now tells remote Cowork or cloud sessions to use a reachable service URL and bearer key. If you cannot operate that endpoint, use the self-contained text-only skill or choose a local workflow.
Current audio and backup bugs block two unattended workflows
Version 0.7.0 added destructive audio watermark removal and per-frame video purification. Open issue 327 identifies a routing gap: Ogg, Opus, and AAC appear in the audio cleaner's extension list, yet the main classifier sends them to the unknown-file refusal. The issue includes an OggS reproduction and says /clean returns 400 for the same inputs. A browser-recorder or Discord pipeline using those formats should wait for a merged fix and release.
Open issue 369 catches a quieter operational problem. clean_file.py --in-place still creates <file>.bak, with no documented --no-backup option in the reported revision. In a static site or deploy directory, that backup may be published and still contain the provenance just removed from the original filename. Keep backups outside scanned and deployed trees, or write cleaned output to a separate staging directory.
October activity is healthy, but a clean result has a narrow meaning
GitHub showed 23,272 stars and 15 combined issues and pull requests on October 3, 2026. The API list separated those into 9 issues and 6 pull requests. The repository was pushed on October 2, while v0.7.0 shipped on September 3. Recent work and the 3 CI workflows show active maintenance, though users still need to check whether a fix has reached the image or tag they deploy.
Use the detailed report as evidence of what a named cleaner attempted and what a named detector observed. Keep the source file. Reopen the output with an independent parser, compare page count or media duration, and review rewritten prose for changed meaning. The passing 96-second suite earns this project a trial. Its own documentation supplies the stopping rule: no tool can promise that an unknown or private detector sees nothing.

