⚡ Auto-agent-skills

Automatically build a reusable personal skills library for your AI agents.

中文版 | English

An MCP server that helps AI agents discover, create, and continuously improve reusable skills — automatically.

✨ Features

Feature	Description
Auto-search	Before a task starts, automatically searches the public skill ecosystem and installs the best matching skill
Auto-review	After a task completes, evaluates whether the solution is worth packaging as a reusable skill
Smart suggestion	Only prompts when a new skill or improvement is genuinely useful — no noise
Continuous improvement	If a used skill underperformed, suggests targeted improvements based on the actual issues
Personal skills library	Maintains a dedicated, version-controlled folder of your own skills at ~/.autoskills/personal-skills/
Multi-agent support	Works with Windsurf, Cursor, Claude Code, and any MCP-compatible agent

---

🚀 Quick Start

1. Install

git clone https://github.com/YOUR_USERNAME/Autoskills.git
cd Autoskills
npm install
npm run build

2. Configure your agent

Add the Autoskills MCP server to your agent's config:

{
  "mcpServers": {
    "auto-agent-skills": {
      "command": "node",
      "args": ["<path-to-autoskills>/dist/index.js"],
      "env": {
        "AUTOSKILLS_DIR": "<path-to-your-home>/.autoskills/personal-skills"
      }
    }
  }
}

AUTOSKILLS_DIR specifies where personal skills created by this MCP are stored.

3. Add Global Rules (Recommended)

Add this global rule (rules/auto-skill-building.md) to your agent to increase the frequency of this MCP's trigger.

4. Start Using It

The agent handles everything automatically:

1. Before a task — if the task looks like it has a relevant public skill, the agent calls search_skill to find and install it
2. During the task — the agent uses the installed skill as a guide
3. After the task — the agent calls review_task to decide whether to create a new skill or improve an existing one

---

🔧 MCP Tools

Tool	When it's called	What it does
search_skill	Before task execution	Searches local global skills (~/.agents/skills/), then falls back to public ecosystem
review_task	After task completion	Evaluates the solution and suggests creating or improving a skill
create_skill	When creating a new skill	Writes a new SKILL.md to the personal library and links it
update_skill	When improving a skill	Updates an existing skill's instructions, description, or metadata
list_skills	On demand	Lists skill counts for global skills and names for personal skills
get_skill	On demand	Reads the full content of a specific skill

---

📚 Personal Skills Library

Skills are stored as Markdown files in ~/.autoskills/personal-skills/:

~/.autoskills/personal-skills/
├── web-scraping/
│   └── SKILL.md
├── docker-setup/
│   └── SKILL.md
└── react-component/
    ├── SKILL.md
    ├── scripts/
    └── references/

Each SKILL.md contains:

---
name: skill-name
description: Short description for matching and triggering
---

# Skill Title

## When to Use
Trigger conditions and applicable scenarios.

## Instructions
Step-by-step workflow for the agent to follow.

Skills are also symlinked to ~/.agents/skills/ so they are globally available to any agent.

---

🛠️ CLI (Optional)

A small CLI for managing skills without an agent:

npx autoskill init <skill-name>      # Create a new skill template
npx autoskill add <path> [-y]        # Add a skill and create symlink
npx autoskill remove <skill-name>    # Remove a skill and its symlink
npx autoskill list                   # List all personal skills

---

💡 Usage Scenarios

Scenario	How it works
Recurring task	Agent finds a matching public skill before starting, uses it, then reviews quality after
Novel solution	Agent completes the task, then suggests packaging the solution as a new personal skill
Skill underperforms	Agent detects issues after using a skill, suggests targeted improvements
Manual skill creation	Use npx autoskill init to scaffold a skill template and fill it in yourself
Sharing skills	Skills are plain Markdown — copy or symlink them across machines or repos

---

📄 License

MIT