___  ____  _____ _   _     ____  _  _____ _     _     ____
                   / _ \|  _ \| ____| \ | |   / ___|| |/ /_ _| |   | |   / ___|
                  | | | | |_) |  _| |  \| |   \___ \| ' / | || |   | |   \___ \
                  | |_| |  __/| |___| |\  |    ___) | . \ | || |___| |___ ___) |
                   \___/|_|   |_____|_| \_|   |____/|_|\_\___|_____|_____|____/

OpenSkills: Run Claude Skills Locally on Your Mac

Anthropic recently announced Skills for Claude - reusable folders with instructions, scripts, and resources that make Claude better at specialized tasks. This tool lets you run these skills entirely on your local machine in a sandboxed environment.

What this means: You can now process your files (documents, spreadsheets, presentations, images) using these specialized skills while keeping all data on your Mac. No uploads, complete privacy.

This tool executes AI-generated code in a truly isolated sandboxed environment on your Mac using Apple's native containers.

Demo

Watch Open-Skills in action with Gemini CLI:

Why Run Skills Locally?

• Privacy: Process sensitive documents, financial data
• Full Control: Skills execute in an isolated container with VM-level isolation
• Compatibility: Works with Claude Desktop, Gemini CLI, Qwen CLI, or any MCP-compatible tool
• Extensibility: Import Anthropic's official skills or create your own custom skills

Quick Start

Prerequisites: Mac with macOS and Apple Silicon (M1/M2/M3/M4/M5), Python 3.10+

git clone https://github.com/BandarLabs/open-skills.git
cd open-skills
chmod +x install.sh
sudo ./install.sh

Installation takes ~2 minutes. The MCP server will be available at http://open-skills.local:8222/mcp

Install required packages (use virtualenv and note the python path):

pip install -r examples/requirements.txt

Setup: Connect Your AI Tool

This MCP server works with any MCP-compatible tool. All execution happens locally on your Mac.

Option 1: Claude Desktop Integration

Configure Claude Desktop to use this MCP server:

1. Copy the example configuration:

   cd examples
   cp claude_desktop/claude_desktop_config.example.json claude_desktop/claude_desktop_config.json

2. Edit the configuration file and replace the placeholder paths:

   • Replace /path/to/your/python with your actual Python path (e.g., /usr/bin/python3 or /opt/homebrew/bin/python3)
   • Replace /path/to/open-skills with the actual path to your cloned repository

   Example after editing:

   {
     "mcpServers": {
       "open-skills": {
         "command": "/opt/homebrew/bin/python3",
         "args": ["/Users/yourname/open-skills/examples/claude_desktop/mcpproxy.py"]
       }
     }
   }

3. Update Claude Desktop configuration:

   • Open Claude Desktop
   • Go to Settings → Developer
   • Add the MCP server configuration
   • Restart Claude Desktop

Option 2: Gemini CLI Configuration

Edit ~/.gemini/settings.json:

{
  "theme": "Default",
  "selectedAuthType": "oauth-personal",
  "mcpServers": {
    "open-skills": {
      "httpUrl": "http://open-skills.local:8222/mcp"
    }
  }
}

For system instructions, replace ~/.gemini/GEMINI.md with GEMINI.md

Option 3: Python OpenAI Agents

Use this server with OpenAI's Python agents library:

1. Set your OpenAI API key:

   export OPENAI_API_KEY="your-openai-api-key-here"

2. Run the client:

   python examples/openai_agents/openai_client.py

Other Supported Tools

• Qwen CLI: Configure similar to Gemini CLI
• Kiro by Amazon: See examples in this repository for configuration
• Any MCP client: Point to http://open-skills.local:8222/mcp

Example Use Cases

Once configured, you can ask your AI to:

• "Create a professional PowerPoint presentation from this markdown outline"
• "Extract all tables from these 10 PDFs and combine into one Excel spreadsheet"
• "Generate ASCII art logo for my project"
• "Fill out this tax form PDF with data from my CSV file"
• "Batch process these 100 images: crop to 16:9 and rotate 90 degrees"

Adding New Skills

You can extend this server with additional skills in two ways:

Option 1: Import Anthropic's Official Skills

Download skills from Anthropic's skills repository and copy to:

~/.open-skills/assets/skills/user/<new-skill-folder>

Available Official Skills:

• Microsoft Word (docx)
• Microsoft PowerPoint (pptx)
• Microsoft Excel (xlsx)
• PDF manipulation
• Image processing
• Slack GIF creator
• And more...

Skills Directory Structure

Here's an example with 4 imported skills:

~/.open-skills/assets/skills/
├── public
│   ├── image-crop-rotate
│   │   ├── scripts
│   │   └── SKILL.md
│   └── pdf-text-replace
│       ├── scripts
│       └── SKILL.md
└── user
    ├── docx
    │   ├── docx-js.md
    │   ├── LICENSE.txt
    │   ├── ooxml
    │   ├── ooxml.md
    │   ├── scripts
    │   └── SKILL.md
    ├── pptx
    │   ├── html2pptx.md
    │   ├── LICENSE.txt
    │   ├── ooxml
    │   ├── ooxml.md
    │   ├── scripts
    │   └── SKILL.md
    ├── slack-gif-creator
    │   ├── core
    │   ├── LICENSE.txt
    │   ├── requirements.txt
    │   ├── SKILL.md
    │   └── templates
    └── xlsx
        ├── LICENSE.txt
        ├── recalc.py
        └── SKILL.md

Option 2: Create Your Own Skills

Create a folder matching the structure shown above. The only mandatory file is SKILL.md. See Anthropic's skills documentation for details.

Quick Method: Ask Claude to generate a skill for you:

"Can you write a skill which creates ASCII art of words?"

Claude will create the skill and offer a ZIP download. Place the ZIP file directly in ~/.open-skills/assets/skills/user (no need to expand).

Manual Method: Create your own skill folder structure:

~/.open-skills/assets/skills/user/my-custom-skill/
├── SKILL.md              # Documentation and usage examples
├── scripts/              # Your Python/bash scripts
│   └── process.py
└── requirements.txt      # (optional) Python dependencies

How Folder Mapping Works

This MCP server provides a compatibility layer that lets you run Claude's skills locally without modification:

Path Translation:

• Claude's path: /mnt/user-data → Local path: /app/uploads
• Skills designed for Claude work locally without any changes

Accessing Your Local Files:

• Place files in ~/.open-skills/assets/outputs on your Mac
• They become available to skills inside the container via volume mounts
• The mapping is automatic - skills can access your files without cloud upload

Skill Structure:

• No changes needed to imported skills
• Original folder hierarchy and file organization remain identical
• Import Claude skills and use them directly

Live Demo: ASCII Art Skill

Here's a real example using Gemini CLI:

> /mcp

Configured MCP servers:

🟢 open-skills - Ready (5 tools)
  Tools:
  - execute_python_code
  - get_skill_file
  - get_skill_info
  - list_skills
  - navigate_and_get_all_visible_text

> can you generate ascii art for my project

✦ I will generate the ASCII art you desire. First, I must survey my available skills.
✓  list_skills (open-skills MCP Server)

✦ I have located a relevant skill: ascii-art. I will now retrieve its instructions.
✓  get_skill_info (open-skills MCP Server) {"skill_name":"ascii-art"}
✦ Your ASCII art is ready:

   ___  ____  _____ _   _     ____  _  _____ _     _     ____
  / _ \|  _ \| ____| \ | |   / ___|| |/ /_ _| |   | |   / ___|
 | | | | |_) |  _| |  \| |   \___ \| ' / | || |   | |   \___ \
 | |_| |  __/| |___| |\  |    ___) | . \ | || |___| |___ ___) |
  \___/|_|   |_____|_| \_|   |____/|_|\_\___|_____|_____|____/




Using: 1 GEMINI.md file | 3 MCP servers (ctrl+t to view)

What happened:

1. AI discovered available skills via list_skills
2. Found the relevant ascii-art skill
3. Retrieved skill instructions with get_skill_info
4. Executed the skill locally in the sandbox
5. Returned results - all without uploading any data to the cloud

Security

Code runs in an isolated container with VM-level isolation. Your host system and files outside the sandbox remain protected.

From @apple/container:

Each container has the isolation properties of a full VM, using a minimal set of core utilities and dynamic libraries to reduce resource utilization and attack surface.

Architecture

This MCP server consists of:

• Sandbox Container: Isolated execution environment with Jupyter kernel
• MCP Server: Handles communication between AI models and the sandbox
• Skills System: Pre-packaged tools for common tasks (PDF manipulation, image processing, etc.)

Available MCP Tools

When connected, this server exposes these tools to your AI:

• execute_python_code - Execute code in the sandbox
• get_skill_file - Read skill files
• get_skill_info - Get skill documentation
• list_skills - List all available skills
• navigate_and_get_all_visible_text - Web scraping with Playwright

Learn More

• GitHub Repository: github.com/BandarLabs/open-skills
• Anthropic Skills: github.com/anthropics/skills
• Skills Documentation: docs.claude.com/skills
• Blog: Building Offline Workspace: instavm.io/blog/building-my-offline-workspace-part-2-skills
• Report Issues: github.com/BandarLabs/open-skills/issues

Contributing

We welcome contributions! If you create useful skills or improve the implementation, please share them with the community.

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.