open-skills

OpenSkills runs Claude-like skills locally on your Mac, enabling private, sandboxed execution of AI-powered workflows without uploading data. The MCP server exposes a single endpoint that you can connect from any MCP-compatible client (Claude Desktop, OpenCode, Gemini CLI, OpenAI agents, Qwen CLI, Kiro, or other MCP clients) at http://open-skills.local:8222/mcp by default. Once connected, you can invoke prebuilt Claude skills (e.g., document processing, image manipulation, or PPTX/Excel workflows) or add your own local skills under your ~/.open-skills/assets/skills directory. The server translates requests from your AI agent into local, isolated tasks that operate inside macOS sandboxed containers, preserving privacy and control over data.

Typical usage involves configuring a client to point at the MCP endpoint, then issuing prompts or tasks that reference the available skills. For example, you can tell Claude Desktop (or another client) to run a PDF text-replace skill, convert a Markdown outline into a PowerPoint-like structure using the local skills, or batch-process a set of images with crop/rotate actions. The setup supports integrating Anthropic’s official skills as well as user-created skills, all accessible through the same MCP interface.

Prerequisites:

• macOS with Apple Silicon (M1/M2/M3/M4/M5)
• Python 3.10 or newer
• Git

1. Clone the repository: git clone https://github.com/BandarLabs/open-skills.git
2. Enter the project directory: cd open-skills
3. Make the install script executable (if provided): chmod +x install.sh
4. Run the installer (installs dependencies and sets up the MCP server): ./install.sh
5. Start the MCP server (if not started by install.sh) using Python module: python -m open_skills.server
6. Verify the server is listening at the default MCP URL: http://open-skills.local:8222/mcp

Notes:

• The installation script installs required Python dependencies listed in examples/requirements.txt.
• If you need to customize paths (e.g., Python binary or project location), modify the environment variables or the mcp_config accordingly.
• Ensure your Mac’s firewall allows incoming connections to port 8222 if you plan to use remote clients.

Tips and common issues:

• If http://open-skills.local:8222/mcp is not reachable, check that the local hostname resolution is working or use 127.0.0.1:8222 as a fallback in your client configuration.
• Use a virtual environment for Python to avoid conflicts with system Python.
• To add new skills, place them under ~/.open-skills/assets/skills/ and ensure each skill has a SKILL.md and a scripts/ entry as described in the repository documentation.
• When integrating with Claude Desktop or Gemini CLI, follow the exact configuration examples provided in the README (paths to Python and the mcpproxy script or equivalent entry points).
• For privacy-sensitive processing, keep all inputs/outputs within ~/.open-skills/assets/outputs/ as described to ensure files are accessible inside the container and not uploaded elsewhere.