🛠️ 用 watermarks-remover 拔除 LLM 隱形字元與中繼資料｜開源好物週報

How could a new project rack up nearly 20,000 stars in just 30 days? Just how amazing is watermarks-remover, the tool that helps developers instantly strip invisible LLM characters? I'm Muyan. In this week's Open Source Weekly, besides this privacy-cleaning powerhouse, we'll also introduce StemDeck, a local mixer that splits music into six tracks, and OpenExecutive, the legendary tool rumored to be built by engineers to replace their bosses. What's the technical magic behind them? I'll tell you all about it in a moment!

This Week's Open Source Picks

1. watermarks-remover: The Ultimate Tool for Cleaning LLM Invisible Watermarks and File Metadata

What it is: This tool specifically addresses a privacy pain point that developers face daily. When we use LLMs to generate code or documentation, invisible Unicode characters are often secretly injected into the content. Alternatively, before publishing files, they might still contain sensitive EXIF and C2PA metadata that shouldn't be leaked. This tool is designed to completely wipe away those traces.
This Week's Buzz: This is a viral new project that exploded within 30 days, racking up 19,801 stars on GitHub. It is licensed under the MIT License and primarily developed in Python.
Key Features:
- Layer A Deterministic Cleaning: Brute-force cleans invisible Unicode, zero-width characters, variation selectors, bidirectional text, and tag characters. The most impressive part is that it purely uses the Python 3.10+ standard library, requiring absolutely no third-party packages.
- Multi-Format Metadata Stripping: Supports over 20 formats including PNG, JPEG, WebP, PDF, DOCX, and MP4. As long as it detects that exiftool, qpdf, or c2patool is installed on the system, it will automatically hook in to clean C2PA and PDF structured data.
- Development Workflow Integration: Can be directly installed as an Agent Skill for Claude Code or Cursor using install_skill.py, hooked up as a pre-commit hook, or even output in SARIF format within CI pipelines for security auditing.

Quick Start: Running an HTTP service locally is the fastest way to get started:

make serve
# 或者用 Docker 跑在本機 8765 埠
docker run --rm -p 127.0.0.1:8765:8765 watermarks-remover

If you just want to process a single file, Layer A requires absolutely no GPU:

python3 service/scripts/clean_file.py draft.md -o draft.cleaned.md

Comparison: Most watermark removal tools on the market require you to upload your files to unknown websites, posing massive privacy risks. In contrast, this tool runs entirely locally and is open-source. Compared to simply using exiftool, it adds text-layer invisible character cleaning and integrates seamlessly into development workflows. However, to be fair, since Anthropic has not disclosed its detector specifications, the project's README honestly admits that there is no guarantee that Layer B's LLM rewriting will 100% bypass statistical watermarking. Furthermore, rewriting might make the original tone sound flat. Its positioning is privacy cleaning, not impersonating human handwriting.
Who it's for: Developers who frequently copy LLM outputs and have extremely high standards for privacy and code cleanliness.
Link: GitHub - guillaumemeyer/watermarks-remover

2. StemDeck: A Free and Fully Local AI Audio Separation Mixer

What it is: This is a local application that helps you split music files into different instrument tracks. It wraps Meta's Demucs into a Digital Audio Workstation (DAW)-style mixer, allowing you to achieve high-quality audio separation right on your own computer without ever uploading your music files to the cloud.
This Week's Buzz: It garnered 244 points (with 65 comments) on Hacker News and has accumulated 3,570 stars on GitHub. It is licensed under Apache-2.0 and primarily developed in JavaScript.
Key Features:
- High-Quality Six-Track Separation: Uses the Demucs htdemucs_6s model to split audio into vocals, drums, bass, guitar, piano, and others. It can also call UVR-MDX-NET Karaoke 2 via audio-separator to further separate lead vocals from backing vocals.
- DAW-Style Mixer and Analysis: Provides track-by-track volume adjustment, mute, solo, and real-time VU meters. The waveform editor supports zooming and looping sections, and it can automatically analyze the song's BPM, key, scale, LUFS loudness, and peak dBFS, finally exporting to a single WAV using FFmpeg.
- Hardware Acceleration Support: Supports directly dragging and dropping local files (MP3, WAV, FLAC, etc.) or pasting YouTube links. The system automatically detects NVIDIA CUDA or Apple Silicon MPS for acceleration, falling back to CPU if no GPU is available.
Quick Start: The easiest way is to go directly to GitHub Releases and download the macOS DMG or Windows ZIP file. On first launch, it will automatically download the Python runtime, FFmpeg, and models (about 170 MB). To self-host using Docker, you can run:
```
docker compose -f build/docker-compose.yml up --build
# 或者直接拉取映像檔
docker pull ghcr.io/stemdeckapp/stemdeck:latest
```
Comparison: Commercial services like Moises or LALAL.AI require subscriptions and uploading files to the cloud, whereas StemDeck is completely free, requires no registration, and keeps your audio files local. The downsides are that it is fixed to six tracks (competitors offer up to ten), runs slower in CPU mode, and has a relatively basic interface. The README even very honestly states, "not looking to compete with commercial products, just doing the core scenario well and stopping there." However, compared to using Demucs directly from the command line, the mixer and waveform interface it adds make it far more user-friendly.
Who it's for: Music creators or audio enthusiasts who need audio track separation, want to avoid paid cloud subscriptions, and value privacy.
Link: GitHub - stemdeckapp/stemdeck

3. OpenExecutive: A Well-Structured Multi-Agent Virtual AI Decision System

What it is: This is a reference implementation of a multi-agent system, claiming to simulate the decision-making and operations of an entire company using 8 specialized AI agents. Although rumors have circulated in the community that this was built by a laid-off engineer as a tool for revenge to replace the CEO, we must objectively state that there is no concrete evidence of this revenge story in public records—it's best treated as water cooler gossip. However, its architecture itself is highly valuable for anyone looking to learn multi-agent orchestration.
This Week's Buzz: The absolute buzzmaker of the week, scoring a high 1,030 points with 717 comments on Hacker News, and accumulating 3,345 stars on GitHub. It is primarily written in Python.
Key Features:
- Eight Specialized Agent Collaboration: Includes CSO, CFO, CHRO, Legal, COO, CMO, CPO, and Board Relations. It uses Claude Sonnet 4.6 as the orchestrator for routing, ultimately converging into a single, consistent persona for external responses.
- Two-Tier RAG and Episodic Memory: Features a built-in MBA-level knowledge base and can ingest user-uploaded company documents into ChromaDB. After each decision, Claude Haiku extracts key points and writes them to SQLite to form long-term memory, allowing subsequent conversations to recall past decisions.
- High-Efficiency Caching and Multi-Interface Support: Caches system prompts, company files, and knowledge indexes separately, claiming a prompt cache hit rate of up to 85%. In addition to the Web UI (Next.js 15), it supports Slack, Telegram, Discord, Email, and CLI, and comes with LLM-as-judge evaluations for 29 scenarios.
Quick Start: Clone the project and set the environment variables to get started:
```
git clone https://github.com/SenteLabsAI/OpenExecutive.git
cd OpenExecutive
cp .env.example .env
# 填入 ANTHROPIC_API_KEY 後執行
make dev
```
Once started, the Web UI will be on port 3000 and the API on 8000. If you want to avoid API costs, you can set LOCAL_MODELS_ENABLED in .env to true to connect to Ollama, LM Studio, or vLLM, though you will lose web search and prompt caching features.
Comparison: Unlike frameworks like CrewAI or AutoGen that require you to build everything from scratch, OpenExecutive is a pre-assembled, ready-to-run vertical application. However, its current maturity should be taken with a grain of salt. The main branch currently has only 39 commits, the built-in scheduler only supports a single instance, and the official docs explicitly state that horizontal scaling is not supported yet. Please treat it as an excellent architectural reference project rather than using it to run a real company just yet.
Who it's for: Developers building AI Agent applications who are looking for reference implementations of multi-agent orchestration, memory layers, and RAG.
Link: GitHub - SenteLabsAI/OpenExecutive

Wrapping Up

From the viral watermarks-remover to the locally run StemDeck, the three picks introduced this week make it clear that "local privacy and security" and "more granular vertical AI applications" remain major development trends. If you are currently struggling with invisible characters generated by LLMs, I highly recommend installing watermarks-remover today to clean things up. I'm Muyan, and this is Open Source Weekly. See you next week!