Skill

You manage and execute multiple skills stored locally with an MCP server that runs on your machine. This server provides unified tools to create, read, update, delete, and run skills, and to execute code that combines multiple skills in a single run. It keeps environment variables and dependencies organized per skill, enabling composable workflows and scalable automation without relying on a single UI.

How to use

You interact with the Skill Management MCP Server through your MCP client (for example Claude Desktop or Cursor). Use the unified CRUD tools to discover, configure, and manage skills in your local skills directory and to execute scripts or inline Python code that can import and compose utilities from multiple skills in one run.

How to install

Prerequisites: ensure you have Python and a tool to install and run MCP servers. You will use uvx to fetch and run the MCP server from PyPI.

{
  "mcpServers": {
    "skill_mcp": {
      "command": "uvx",
      "args": [
        "--from",
        "skill-mcp",
        "skill-mcp-server"
      ]
    }
  }
}

Configuration and quick start

Configure your MCP client to connect to the local MCP server using the provided command. The server is installed and run via uvx from PyPI, so you don’t need to place a server file locally.

Example to add the MCP server in your client configuration (standard stdio-based approach):

{
  "mcpServers": {
    "skill_mcp": {
      "command": "uvx",
      "args": [
        "--from",
        "skill-mcp",
        "skill-mcp-server"
      ]
    }
  }
}

Testing the setup

Restart your MCP client after configuring the server. In a new conversation, issue a request like:

List all available skills

Directory and runtime basics

Skills are stored in a local directory, typically under your home path. A typical layout is shown here with a sample skill named example-skill and its components.

~/.skill-mcp/
└── skills/
    └── example-skill/
        ├── SKILL.md
        ├── .env
        ├── scripts/
        ├── references/
        └── assets/

Security and best practices

Keep sensitive values in per-skill .env files with restrictive permissions. The server reads environment variables at execution time and merges env vars from referenced skills. Do not expose secret values in logs or in code.

Enable a safe execution boundary by using the provided 30-second timeout for script runs and ensuring scripts only access intended resources.

Troubleshooting

If the MCP server cannot be found, verify that uv is installed and available on your PATH, and ensure the skills directory path is correct. If you encounter permission errors, adjust file permissions on the MCP directories and .env files as needed.

For scripts failing to execute, confirm execute permissions, interpreter availability, and that required environment variables are present.

Advanced: Multi-skill inline execution

You can run Python code that imports utilities from multiple skills in one execution. This unifies tools from different skills into a single runtime, loading dependencies and env vars from all referenced skills automatically.

Example usage concept: import modules from different skills and run a combined workflow in a single code block. The environment loads dependencies via PEP 723 inline metadata where supported and aggregates them across skills.

Available tools

skill_crud

Unified CRUD tool for managing skills: list, get, create, delete, validate, list_templates

skill_files_crud

Unified file CRUD tool for skill files: read, create, update, delete (supports bulk operations)

skill_env_crud

Unified environment variable CRUD tool: read, set, delete, clear

run_skill_script

Execute scripts with automatic dependency detection and environment injection (PEP 723 support)

execute_python_code

Execute Python code directly, enabling cross-skill imports and automatic dependency aggregation