Skill Management MCP Server

A Model Context Protocol (MCP) server that enables Claude to manage skills stored in ~/.skill-mcp/skills. This system allows Claude to create, edit, run, and manage skills programmatically, including execution of skill scripts with environment variables.

Quick Status

Status: ✅ Production Ready Test Coverage: 86% (145/145 tests passing) Deployed: October 18, 2025 Architecture: 22-module modular Python package with unified CRUD architecture

Overview

TL;DR: Write Python code that unifies multiple skills in one execution - follows Anthropic's MCP pattern for 98.7% more efficient agents.

This project consists of two main components:

1. MCP Server (src/skill_mcp/server.py) - A Python package providing 5 unified CRUD tools for skill management

2. Skills Directory (~/.skill-mcp/skills/) - Where you store and manage your skills

Key Advantages

🚀 Unified Multi-Skill Execution (Code Execution with MCP)

Build once, compose everywhere - Execute Python code that seamlessly combines multiple skills in a single run:

What makes this powerful:

• ✅ Context-efficient - Dependencies and env vars auto-aggregated from all referenced skills

• ✅ Composable - Mix and match utilities from any skill like building blocks

• ✅ No redundancy - Declare PEP 723 dependencies once in library skills, reuse everywhere

• ✅ Progressive disclosure - Load only the skills you need, when you need them

• ✅ Follows Anthropic's MCP pattern - Code execution with MCP for efficient agents

Efficiency gains:

• 📉 98.7% fewer tokens when discovering tools progressively vs loading all upfront

• 🔄 Intermediate results stay in code - Process large datasets without bloating context

• ⚡ Single execution - Complex multi-step workflows in one code block instead of chained tool calls

This aligns with Anthropic's research showing agents scale better by writing code to call tools rather than making direct tool calls for each operation.

🔓 Not Locked to Claude UI

Unlike the Claude interface, this system uses the Model Context Protocol (MCP), which is:

• ✅ Universal - Works with Claude Desktop, claude.ai, Cursor, and any MCP-compatible client

• ✅ Not tied to Claude - Same skills work everywhere MCP is supported

• ✅ Future-proof - Not dependent on Claude's ecosystem or policy changes

• ✅ Local-first - Full control over your skills and data

🎯 Use Skills Everywhere

Your skills can run in:

• Cursor - IDE integration with MCP support

• Claude Desktop - Native app with MCP access

• claude.ai - Web interface with MCP support

• Any MCP client - Growing ecosystem of compatible applications

📦 Independent & Modular

• ✅ Each skill is self-contained with its own files, scripts, and environment

• ✅ No dependency on proprietary Claude features

• ✅ Can be versioned, shared, and reused across projects

• ✅ Standard MCP protocol ensures compatibility

🔄 Share Skills Across All MCP Clients

• ✅ One skill directory, multiple clients - Create once, use everywhere

• ✅ Same skills in Cursor and Claude - No duplication needed

• ✅ Seamless switching - Move between tools without reconfiguring

• ✅ Consistent experience - Skills work identically across all MCP clients

• ✅ Centralized management - Update skills in one place, available everywhere

🤖 LLM-Managed Skills (No Manual Copy-Paste)

Instead of manually copying, zipping, and uploading files:

Key Benefits:

• ✅ No manual file operations - LLM handles creation, editing, deletion

• ✅ Instant changes - No upload/download/reload cycles

• ✅ Full version control - Skills are regular files, can use git

• ✅ Easy modification - LLM can edit scripts on the fly

• ✅ Testable - LLM can create and run scripts immediately

• ✅ Collaborative - Teams can develop skills together via MCP

Features

Skill Management

• ✅ List all available skills

• ✅ Browse skill files and directory structure

• ✅ Read skill files (SKILL.md, scripts, references, assets)

• ✅ Create new skill files and directories

• ✅ Update existing skill files

• ✅ Delete skill files

Script Execution

• ✅ Run Python, Bash, and other executable scripts

• ✅ Automatic dependency management for Python scripts using uv inline metadata (PEP 723)

• ✅ Automatic environment variable injection from secrets

• ✅ Command-line argument support

• ✅ Custom working directory support

• ✅ Capture stdout and stderr

• ✅ 30-second timeout for safety

Direct Python Execution - Multi-Skill Unification 🚀

• ✅ UNIFY MULTIPLE SKILLS in one execution - Combine utilities from different skills seamlessly

• ✅ Execute Python code directly without creating script files

• ✅ Cross-skill imports - Import modules from ANY skill as reusable libraries

• ✅ Automatic dependency aggregation - Dependencies from ALL imported skills auto-merged

• ✅ Environment variable loading - .env files from ALL referenced skills auto-loaded

• ✅ PEP 723 support - Inline dependency declarations in code

• ✅ 98.7% more efficient - Follows Anthropic's recommended MCP pattern for scalable agents

• ✅ Perfect for multi-skill workflows, quick experiments, data analysis, and complex pipelines

Environment Variables

• ✅ List environment variable keys (secure - no values shown)

• ✅ Set or update environment variables per skill

• ✅ Persistent storage in per-skill .env files

• ✅ Automatic injection into script execution

Directory Structure

Note: The MCP server is installed via uvx from PyPI and runs automatically. No local server file needed!

Quick Start

1. Install uv

This project uses uv for fast, reliable Python package management.

2. Configure Your MCP Client

Add the MCP server to your configuration. The server will be automatically downloaded and run via uvx from PyPI.

Claude Desktop - Edit the config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• Linux: ~/.config/Claude/claude_desktop_config.json

Cursor - Edit the config file:

• macOS: ~/.cursor/mcp.json

• Windows: %USERPROFILE%\.cursor\mcp.json

• Linux: ~/.cursor/mcp.json

That's it! No installation needed - uvx will automatically download and run the latest version from PyPI.

3. Restart Your MCP Client

Restart Claude Desktop or Cursor to load the MCP server.

4. Test It

In a new conversation:

Claude should use the skill-mcp tools to show skills in ~/.skill-mcp/skills/.

Common uv Commands

For development in this repository:

Note: uv automatically creates and manages .venv/ - no need to manually create virtual environments!

Script Dependencies (PEP 723)

✅ BOTH

Python scripts and code can declare their own dependencies using uv's inline metadata. The server automatically detects this and uses uv run to handle dependencies:

Benefits:

• ✅ No manual dependency installation needed

• ✅ Each script/code execution has isolated dependencies

• ✅ Works automatically with both run_skill_script and execute_python_code

• ✅ Version pinning ensures reproducibility

• ✅ execute_python_code ALSO aggregates dependencies from skill imports!

How it works with

1. You add inline metadata to your Python script file

2. When the script runs via run_skill_script, the server detects the metadata

3. uv automatically creates an isolated environment and installs dependencies

4. The script runs with access to those dependencies

5. No manual pip install or virtual environment management needed!

How it works with

1. Include PEP 723 metadata directly in your code string

2. The server automatically detects the metadata

3. uv creates an isolated environment and installs dependencies

4. Your code runs with access to those dependencies

5. BONUS: If you import from skill files, their PEP 723 dependencies are automatically aggregated too!

Example: See example-skill/scripts/fetch_data.py for a working example.

Testing locally:

Direct Python Code Execution - Unify Multiple Skills in One Run

The execute_python_code tool allows you to run Python code that combines multiple skills in a single execution. This is perfect for:

• 🔄 Multi-skill workflows - Import and compose utilities from different skills

• 🧪 Quick experiments - Test code without creating files

• 📊 Data analysis - Process data using libraries from multiple skills

• 🏗️ Building on reusable skill libraries - Create specialized utilities once, use everywhere

Key insight from Anthropic's research: Agents scale better by writing code to call tools instead of making direct tool calls. This approach reduces context usage by up to 98.7% and enables more efficient workflows.

Basic Usage

Cross-Skill Imports - Unifying Multiple Skills

The power of composition - Create utility skills once and combine them in endless ways:

Real-world example: Process sales data by unifying calculator, data-processor, and CRM skills:

Step 1: Create a calculator skill with reusable modules

Step 2: Create data-processor skill utilities

Step 3: Unify both skills in one execution!

What just happened:

• ✅ Two skills unified - calculator + data-processor in one execution

• ✅ Zero redundancy - pandas dependency declared once in csv_parser.py, auto-included

• ✅ Composable - Mix and match any skills like LEGO blocks

• ✅ Context-efficient - Only loaded the specific modules needed

Automatic Dependency Aggregation

When you import from skill modules that have PEP 723 dependencies, they're automatically included:

Library skill with dependencies:

Your code - NO need to redeclare requests!

Environment Variables from Referenced Skills

When you import from a skill, its environment variables are automatically loaded:

Skill with API credentials:

Your code - env vars automatically available:

Benefits:

• ✅ No need to manually load .env files

• ✅ Each skill's secrets stay isolated

• ✅ Multiple skills' env vars are merged automatically

• ✅ Later skills override earlier ones if there are conflicts

Use Cases

• 🔄 Multi-skill workflows - THE KILLER FEATURE - Unify utilities from multiple skills in one execution

  • Example: Combine API client + data parser + analytics calculator in single run

  • Example: Chain together scraper + NLP processor + notification sender

  • Example: Merge CRM data + payment processor + reporting tools

• ✅ Quick data analysis - Run pandas/numpy code without creating files

• ✅ API testing - Test HTTP requests with inline dependencies

• ✅ Reusable libraries - Build once, import everywhere

• ✅ Rapid prototyping - Experiment with code before committing to files

• ✅ Complex pipelines - Build multi-stage data processing in one code block

Comparison: run_skill_script vs execute_python_code

Both tools support PEP 723, but have different use cases:

Key Insight:

• Use run_skill_script when you have a script file ready to execute

• Use execute_python_code when you want to UNIFY MULTIPLE SKILLS in one execution - This is the recommended approach per Anthropic's MCP research for building efficient, scalable agents

Usage Examples

Creating a New Skill

Managing Environment Variables

Running Skill Scripts

Modifying Existing Skills

Available MCP Tools

The server provides these unified CRUD tools to Claude:

Key Benefits of CRUD Architecture:

• ✅ Reduced context window usage - 5 tools instead of 9+

• ✅ Consistent operation patterns - All tools follow the same CRUD model

• ✅ Bulk operations - Create/update/delete multiple files atomically

• ✅ Better error handling - Unified error responses across all operations

Security Features

Path Validation

• All file paths are validated to prevent directory traversal attacks

• Paths with ".." or starting with "/" are rejected

• All operations are confined to the skill directory

Environment Variables

• Variable values are never exposed when listing

• Stored in per-skill .env files

• File permissions should be restricted (chmod 600 on each .env)

Script Execution

• 30-second timeout prevents infinite loops

• Scripts run with user's permissions (not elevated)

• Output size limits prevent memory issues

• Capture both stdout and stderr for debugging

Troubleshooting

"MCP server not found"

• Check that uv is in your PATH: which uv (or where uv on Windows)

• Verify the path to .skill-mcp directory is correct and absolute

• Test dependencies: cd ~/.skill-mcp && uv run python -c "import mcp; print('OK')"

• Ensure pyproject.toml exists in ~/.skill-mcp/

"Permission denied" errors

Scripts failing to execute

• Check script has execute permissions

• Verify interpreter (python3, bash) is in PATH

• Use list_env_keys to check required variables are set

• Check stderr output from run_skill_script

Environment variables not working

• Verify they're set: use read_skill_env for the skill

• Check the .env file exists: cat ~/.skill-mcp/skills/<skill-name>/.env

• Ensure your script is reading from os.environ

Advanced: CRUD Tool Operations

All MCP tools follow a unified CRUD architecture with detailed descriptions:

skill_crud Operations

• list - List all skills with descriptions, paths, and validation status (supports text/regex search)

• get - Get comprehensive skill information: SKILL.md content, all files, scripts, environment variables

• create - Create new skill from template (basic, python, bash, nodejs)

• delete - Delete a skill directory (requires confirmation)

• validate - Validate skill structure and get diagnostics

• list_templates - List all available skill templates with descriptions

skill_files_crud Operations

• read - Read one or multiple files in a skill directory (supports bulk reads)

• create - Create one or more files (auto-creates parent directories, supports atomic bulk creation)

• update - Update one or more existing files (supports bulk updates)

• delete - Delete a file permanently (path-traversal protected, SKILL.md cannot be deleted)

skill_env_crud Operations

• read - List environment variable keys for a skill (values hidden for security)

• set - Set one or more environment variables (merges with existing)

• delete - Delete one or more environment variables

• clear - Clear all environment variables for a skill

Script Execution

• run_skill_script - Execute scripts with automatic PEP 723 dependency detection and environment variable injection

• execute_python_code - Execute Python code directly without files (supports PEP 723 dependencies and cross-skill imports)

Advanced Configuration

Custom Skills Directory

The skills directory can be customized using the SKILL_MCP_DIR environment variable. If not set, it defaults to ~/.skill-mcp/skills.

Setting via environment variable (recommended):

In MCP client configuration:

For Claude Desktop or Cursor, add the environment variable to your MCP config:

Notes:

• The directory will be created automatically if it doesn't exist

• Use absolute paths for the custom directory

• All skills will be stored in the configured directory

• No global secrets file; env vars are per-skill .env files

Resource Limits

Resource limits are defined in src/skill_mcp/core/config.py:

To modify these limits, you'll need to fork the repository and adjust the constants in the config file.

Architecture & Implementation

Package Structure

What's New

Unified CRUD Architecture:

• ✅ 3 unified CRUD tools instead of 9+ individual tools (skill_crud, skill_files_crud, skill_env_crud)

• ✅ Bulk operations - Create/update/delete multiple files atomically

• ✅ Consistent patterns - All tools follow the same operation-based model

• ✅ Better error handling - Unified error responses across all operations

Direct Python Execution (Multi-Skill Unification):

• 🚀 execute_python_code - UNIFY MULTIPLE SKILLS in one execution (Anthropic's recommended MCP pattern)

• ✅ Cross-skill imports - Import modules from ANY skill as reusable libraries

• ✅ Automatic dependency aggregation - Dependencies from ALL imported skills auto-merged

• ✅ Automatic environment loading - .env files from ALL referenced skills auto-loaded

• ✅ PEP 723 support - Inline dependency declarations

• 📉 98.7% token reduction - Load skills progressively instead of all upfront

Enhanced Features:

• ✅ Skill templates - Create skills from templates (basic, python, bash, nodejs)

• ✅ Template discovery - List all available templates with descriptions

• ✅ Skill validation - Validate skill structure and get diagnostics

• ✅ Search capabilities - Search skills by name/description with text or regex

• ✅ Namespaced paths - File paths shown as "skill_name:file.py" for clarity

• ✅ Configurable skills directory - Use SKILL_MCP_DIR environment variable

Test Results

Unit Tests: 145/145 Passing ✅

Coverage: 86% (959/1120 statements covered)

Comprehensive test coverage across all modules:

Test Organization:

• ✅ CRUD operations: Comprehensive tests for all operations (create, read, update, delete)

• ✅ Bulk operations: Atomic transaction tests for file operations

• ✅ Template system: Template discovery, validation, and creation

• ✅ Path security: Directory traversal prevention and validation

• ✅ PEP 723 support: Dependency detection and aggregation

• ✅ Integration tests: Full MCP server workflow testing

Manual Tests: All Passed ✅

• ✅ List skills with YAML descriptions and search functionality

• ✅ Get comprehensive skill details with SKILL.md content

• ✅ Create skills from templates (basic, python, bash, nodejs)

• ✅ Read/create/update/delete files (single and bulk)

• ✅ Read/set/delete/clear environment variables

• ✅ Execute scripts with auto-dependencies (PEP 723)

• ✅ Execute Python code directly with cross-skill imports

• ✅ Dependency aggregation from imported skill modules

• ✅ Environment variable loading from referenced skills

Verification Checklist

• ✅ Server imports successfully

• ✅ All 5 unified CRUD tools registered and callable

• ✅ 145/145 unit tests passing (86% coverage)

• ✅ All manual tests passing

• ✅ MCP client configuration working (Claude Desktop, Cursor)

• ✅ Package deployed to PyPI and active

• ✅ Scripts execute successfully with PEP 723 dependencies

• ✅ File operations working (including bulk operations)

• ✅ Environment variables working (CRUD operations)

• ✅ Template system working (create, list, validate)

• ✅ Direct Python execution working with cross-skill imports

• ✅ Backward compatible with existing skills

Best Practices

Skill Development

• Follow the standard skill structure (SKILL.md, scripts/, references/, assets/)

• Keep SKILL.md concise and focused

• Use progressive disclosure (split large docs into references)

• Test scripts immediately after creation

Environment Variables

• Use descriptive names (API_KEY, DATABASE_URL)

• Never log or print sensitive values

• Set permissions on .env files: chmod 600 ~/.skill-mcp/skills/<skill-name>/.env

Script Development

• Use meaningful exit codes (0 = success)

• Print helpful messages to stdout

• Print errors to stderr

• Include error handling

• For Python scripts with dependencies: Use inline metadata (PEP 723)

  # /// script # dependencies = [ # "package-name>=version", # ] # ///

• Scripts without metadata use the system Python interpreter

• Scripts with metadata automatically get isolated environments via uv

🔐 Managing Sensitive Secrets Safely

To prevent LLMs from accessing your sensitive credentials:

✅ RECOMMENDED: Update .env files directly on the file system

Why this is important:

• ✅ LLMs never see your sensitive values

• ✅ Secrets stay on your system only

• ✅ No risk of credentials appearing in logs or outputs

• ✅ Full control over sensitive data

• ✅ Can be used with git-secret or similar tools for versioning

Workflow:

1. Claude creates the skill structure and scripts

2. You manually add sensitive values to .env files

3. Claude can read the .env keys (without seeing values) and use them

4. Scripts access secrets via environment variables at runtime

Example:

❌ NEVER DO:

• ❌ Tell Claude your actual API keys or passwords

• ❌ Ask Claude to set environment variables with sensitive values

• ❌ Store secrets in SKILL.md or other tracked files

• ❌ Use update_skill_env tool with real secrets (only for non-sensitive config)

✅ DO:

• ✅ Update .env files manually on your system

• ✅ Keep .env files in .gitignore

• ✅ Use chmod 600 to restrict file access

• ✅ Tell Claude only the variable names (e.g., "the API key is in API_KEY")

• ✅ Keep secrets completely separate from LLM interactions

⚠️ Important: Verify LLM-Generated Code

When Claude or other LLMs create or modify skills and scripts using this MCP system, always verify the generated code before running it in production:

Security Considerations

• ⚠️ Always review generated code - LLMs can make mistakes or generate suboptimal code

• ⚠️ Check for security issues - Look for hardcoded credentials, unsafe operations, or vulnerabilities

• ⚠️ Test thoroughly - Run scripts in isolated environments first

• ⚠️ Validate permissions - Ensure scripts have appropriate file and system permissions

• ⚠️ Monitor dependencies - Review any external packages installed via PEP 723

Best Practices for LLM-Generated Skills

1. Review before execution - Always read through generated scripts

2. Test in isolation - Run in a safe environment before production use

3. Use version control - Track all changes with git for audit trails

4. Implement error handling - Add robust error handling and logging

5. Set resource limits - Use timeouts and resource constraints

6. Run with minimal permissions - Don't run skills as root or with elevated privileges

7. Validate inputs - Sanitize any user-provided data

8. Audit logs - Review what scripts actually do and track their execution

Common Things to Check

• ❌ Hardcoded API keys, passwords, or tokens

• ❌ Unsafe file operations or path traversal risks

• ❌ Unvalidated external commands or shell injection risks

• ❌ Missing error handling or edge cases

• ❌ Resource-intensive operations without limits

• ❌ Unsafe deserialization (eval, pickle, etc.)

• ❌ Excessive permissions requested

• ❌ Untrustworthy external dependencies

When in Doubt

• Ask Claude/LLM to explain the code

• Have another person review critical code

• Use linters and security scanning tools

• Run in containers or VMs for isolation

• Start with read-only operations before destructive ones

Remember: LLM-generated code is a starting point. Your verification and review are essential for security and reliability.

Installation from PyPI

To install the package globally (optional):

Or use uvx to run without installation (recommended):

Development Setup

If you want to contribute or run from source:

To use your local development version in your MCP client config:

License

MIT License

Copyright (c) 2025

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Contributing

This is a custom tool for personal use. Feel free to fork and adapt for your needs.

Support

For setup issues or questions, refer to:

• Claude's MCP documentation at https://modelcontextprotocol.io

• The MCP Python SDK docs at https://github.com/modelcontextprotocol/python-sdk