Agent Skills

Enable any AI Agent to gain Skills capability with one click, driving Agent self-evolution

📖 中文文档 (Chinese Documentation)

Vision

We believe AI Agents should have the ability to self-evolve—not just execute tasks, but also learn new skills and create new tools.

Skills are the core mechanism to realize this vision:

• Agents can learn pre-packaged professional knowledge and scripts at any time
• Agents can create entirely new skills as needed
• Skills exist as independent modules that can be shared and reused

Agent Skills enables any AI Agent to gain this capability with one click:

• MCP Protocol Support: Compatible with Claude Desktop, Cursor, etc.
• LangChain Middleware Support: Native integration into your Agent applications
• Docker Isolated Execution: Secure, reliable, and ready to use out of the box
• Progressive Disclosure: Lightweight loading, on-demand reading

┌─────────────────────────────────────────────────────────────┐
│                      Your AI Agent                          │
│                                                             │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐         │
│  │ LangChain   │  │ Claude      │  │ Custom      │         │
│  │ Agent       │  │ Desktop     │  │ Agent       │         │
│  └──────┬──────┘  └──────┬──────┘  └──────┬──────┘         │
│         │                │                │                 │
│         └────────────────┼────────────────┘                 │
│                          │                                  │
│                          ▼                                  │
│              ┌───────────────────────┐                      │
│              │    Agent Skills       │                      │
│              │  ┌─────┐ ┌─────────┐  │                      │
│              │  │ MCP │ │Middleware│ │                      │
│              │  └─────┘ └─────────┘  │                      │
│              └───────────────────────┘                      │
│                          │                                  │
│                          ▼                                  │
│              ┌───────────────────────┐                      │
│              │   Skills Ecosystem    │                      │
│              │  PDF | Code Review    │                      │
│              │  Data Analysis | ...  │                      │
│              └───────────────────────┘                      │
└─────────────────────────────────────────────────────────────┘

Features

• Unified Toolset: 6 skills_* tools with atomic functionality, easy to understand
• Docker Isolated Execution: Run commands in containers with pre-installed tools and libraries
• Dual Integration Options:
• MCP Protocol: Compatible with Claude Desktop, Cursor, etc.
• LangChain Middleware: Native Python integration with lower latency
• Progressive Disclosure: Skill metadata pre-loaded, full content read on demand
• Meta-skill Auto-copy: Automatically get skill-creator when using custom skills directory

Built-in Skills

In addition to the core framework capabilities, we provide several carefully designed skills ready to use:

🌟 We look forward to you creating more interesting skills! Using the skill-creator meta-skill, you can easily package your professional knowledge and tools. If you create useful skills, feel free to contribute to the community.

Quick Start

1. Build Docker Image

docker build -t agent-skills:latest -f docker_config/Dockerfile .

2. Choose Integration Method

Option A: Claude Desktop / Cursor (MCP)

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "agent-skills": {
      "command": "docker",
      "args": ["run", "-i", "--rm",
               "-v", "~/.agent-skills/skills:/skills",
               "-v", "/Users:/Users",
               "agent-skills:latest"]
    }
  }
}

The two mounts serve these purposes:

• ~/.agent-skills/skills:/skills - Skills directory (required), modify to your preferred storage location
• /Users:/Users - Host file access (optional, for scripts to read/write external files)

💡 Tip: To get the same effect as Middleware (which can inject system prompts) when using MCP, add a .cursor/rules/python/skills_prompt.mdc file to your project root. This helps guide Cursor to use skills effectively. See MCP Integration for details.

💡 On Linux, change /Users:/Users to /home:/home

Option B: LangChain Application (Middleware)

from agent_skills.core.middleware import SkillsMiddleware
from deepagents import create_deep_agent

# Configure skills_dir and host directory mount
middleware = SkillsMiddleware(
    skills_dir="/path/to/skills",
    host_mount="/Users:/Users",  # Optional, for scripts to access external files
)

agent = create_deep_agent(
    tools=[],
    system_prompt="You are a helpful assistant.",
    middleware=middleware.get_middlewares(),
)

Documentation

Development

# Install dependencies
uv sync

# Run tests
uv run pytest tests/ -v

# Start MCP Server locally
uv run agent-skills-server

License

Apache 2.0