@agiflowai/openskill

Model Context Protocol (MCP) server that enables Anthropic Skills to work with any coding agent and client that supports MCP servers. Provides skill management and sandboxed execution capabilities.

Features

• Universal Compatibility: Makes Anthropic Skills work with any MCP-compatible AI coding agent (Cline, Roo Code, Continue, Zed, etc.)
• Skill Management: Read and execute skills from .claude/skills/ directory
• Sandboxed Execution: Execute bash commands in isolated Docker containers for security
• Volume Mounting: Mount host directories into containers for file access
• MCP Tools: get-skill and use-skill tools for seamless integration
• CLI Commands: Direct command-line access to all functionality
• Custom Images: Support for custom Docker images tailored to your needs
• Pre-configured Environment: Python (uv, pip, poetry, pipenv, Pillow) and Node.js tooling out of the box

MCP Tools

get-skill

Retrieves skill information from the skills directory. You can use anthopic's skills in github and put that in .claude/skills or your custom skill folder and that should work.

Input:

• command (string): Name of the skill to retrieve

Example:

{
  "command": "canvas-design"
}

use-skill

Executes bash commands in a sandboxed Docker environment for a specific skill.

Input:

• skillName (string): Name of the skill to execute
• bash (string): Bash command to execute

Example:

{
  "skillName": "canvas-design",
  "bash": "python generate.py output.pdf"
}

CLI Commands

mcp-serve

Start the MCP server with stdio transport.

node dist/cli.js mcp-serve [options]

Options:

• -t, --type <type>: Transport type (default: "stdio")
• --timeout <ms>: Default timeout for use-skill command (default: 30000)
• --workdir <path>: Default working directory in container (default: "/workspace")
• --mount <path>: Host path to mount into container /workspace
• --skills-path <path>: Path to skills directory (default: ".claude/skills")
• --container-name <name>: Custom Docker container name (allows reusing one container for multiple skills)
• --image <name>: Custom Docker image name to use instead of building default openskill-http image
• --disable-tools <tools>: Comma-separated list of tools to disable
• --no-prewarm: Disable Docker image prewarming on server startup

Example:

# Start with custom skills path
node dist/cli.js mcp-serve --skills-path ./my-skills

# Start with custom Docker image
node dist/cli.js mcp-serve --image my-custom-image:latest

# Start with custom container name (reuse same container for all skills)
node dist/cli.js mcp-serve --container-name shared-skill-container

use-skill

Execute a skill directly from the command line.

node dist/cli.js use-skill <skill-name> <bash-command> [options]

Options:

• -t, --timeout <ms>: Command timeout (default: 30000)
• -w, --workdir <path>: Working directory (default: "/workspace")
• -m, --mount <path>: Host path to mount into container

Example:

# Execute Python script in canvas-design skill
node dist/cli.js use-skill canvas-design "python generate.py output.pdf" \
  --mount /Users/username/projects

doctor

Check if Docker is installed and accessible. Validates your environment for running OpenSkill.

node dist/cli.js doctor

Checks:

• Docker installation and version
• Docker daemon status

Example Output:

Running OpenSkill environment checks...

Checking Docker installation...
✓ Docker is installed: Docker version 28.3.2, build 578ccf6

Checking Docker daemon...
✓ Docker daemon is running

✓ All checks passed!
  Your environment is ready to use OpenSkill.

http-serve

Start the HTTP server (used by Docker containers). When you create your own dockerfile, simply use npx @agiflowai/openskill http-serve as your entry point.

node dist/cli.js http-serve [options]

Options:

• -p, --port <port>: Port to listen on (default: 3000)
• -h, --host <host>: Host to bind to (default: "localhost")

Usage with MCP Clients

OpenSkill can be used with any MCP-compatible client. Below are basic configuration examples for popular tools.

Cline

{
  "mcpServers": {
    "openskill-mcp": {
      "disabled": false,
      "command": "openskill",
      "args": ["mcp-serve"]
    }
  }
}

Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "openskill-mcp": {
      "command": "openskill",
      "args": ["mcp-serve"]
    }
  }
}

Windsurf

{
  "mcpServers": {
    "openskill-mcp": {
      "command": "openskill",
      "args": ["mcp-serve"]
    }
  }
}

VS Code

"mcp": {
  "servers": {
    "openskill-mcp": {
      "type": "stdio",
      "command": "openskill",
      "args": ["mcp-serve"]
    }
  }
}

Zed

{
  "context_servers": {
    "openskill-mcp": {
      "source": "custom",
      "command": "openskill",
      "args": ["mcp-serve"]
    }
  }
}

Roo Code

{
  "mcpServers": {
    "openskill-mcp": {
      "command": "openskill",
      "args": ["mcp-serve"]
    }
  }
}

Continue

{
  "mcpServers": {
    "openskill-mcp": {
      "command": "openskill",
      "args": ["mcp-serve"]
    }
  }
}

Claude Code

{
  "mcpServers": {
    "openskill-mcp": {
      "command": "openskill",
      "args": ["mcp-serve", "--disable-tools", "get-skill"]
    }
  }
}

Note: For advanced configuration options (mount paths, container names, custom Docker images, skills paths, timeouts, etc.), see the CLI Commands section above.

Docker Environment

The sandbox Docker image includes:

Node.js/JavaScript:

• Node.js v20
• npm

Python:

• Python 3.12
• pip
• uv (fast package installer)
• poetry (dependency management)
• pipenv (virtual environment manager)
• Pillow (PIL - image processing library)

System Tools:

• bash
• git
• curl, wget
• Build tools (gcc, g++, make)

Building the Docker Image

cd packages/openskill
docker build -t openskill-http .

Development

# Install dependencies
pnpm install

# Build the project
pnpm build

# Run type checking
pnpm typecheck

# Run tests
pnpm test

# Development mode (watch)
pnpm dev

Skill Structure

Skills should be organized in the skills directory (default: .claude/skills/):

.claude/skills/
├── canvas-design/
│   ├── SKILL.md          # Skill definition with frontmatter
│   ├── generate.py       # Skill implementation
│   └── templates/        # Additional resources
└── another-skill/
    └── SKILL.md

SKILL.md Format:

---
name: canvas-design
description: Create beautiful visual art in .png and .pdf documents
license: user
---

# Canvas Design Skill

Instructions for the AI on how to use this skill...

Architecture

• src/tools/: MCP tool implementations (GetSkillTool, UseSkillTool)
• src/services/: Business logic (SkillService, SandboxService)
• src/commands/: CLI commands (mcp-serve, use-skill, http-serve)
• src/server/: MCP server setup and configuration
• src/transports/: MCP transport handlers (stdio)
• Dockerfile: Docker image with Python and Node.js tooling

License

MIT