SkillNet

What is SkillNet?

SkillNet is an open-source research and engineering infrastructure platform developed by the Zhejiang University NLP Lab (zjunlp) and OpenKG (presented in arXiv:2603.04448). Designed to solve the fragmentation, unverified quality, and manual discovery challenges of reusable AI agent capabilities, SkillNet indexes over 500,000+ GitHub skills. It provides an end-to-end lifecycle system covering Discovery, Installation, Creation, Evaluation, Analysis, and Dynamic Task Routing across coding assistants and autonomous agent hosts like Claude Code, Cursor, OpenClaw, and JiuwenClaw.

Core Architectural Capabilities

• Full Agent Skill Lifecycle:
  • Discovery: High-speed keyword and vector semantic search across 500K+ deduplicated GitHub skills categorized by domain.
  • Installation: Automated verification and downloading of skill directory structures into local agent environments with a single command.
  • Creation: Prompt-driven and trace-driven skill package synthesis, extracting reusable patterns from codebases, execution logs, and documentation.
  • Evaluation: Multi-dimensional scoring assessing skills on safety, structural completeness, executability, maintainability, and token cost awareness.
  • Analysis & Scenario Graphs: Analyzes local skill libraries to infer directional relationships (such as compose_with and similar_to) with source line citations.
  • Routing (SkillNet-Fabric): Dynamically selects the optimal subset of local skills for a given complex task, generating an on-the-fly task Wiki with explicit selection rationales.
• SkillNet-Gym Benchmarks: An executable evaluation suite testing agent skill construction, retrieval, and composition across interactive environments like ALFWorld, WebShop, and ScienceWorld.
• Local Browser UI (skillnet ui): A lightweight local web interface running at http://127.0.0.1:8765 allowing developers to explore skill trees, inspect scenario graphs, and review source lines without external API keys or Node.js runtimes.
• Native Model Context Protocol (MCP) Support: Official integration via skillnet-mcp (maintained by CycleChain) connects SkillNet directly to Claude Code, Cursor, and any MCP-compatible agent client.
• Zero-Auth Public REST API: Free, public search endpoints at http://api-skillnet.openkg.cn/v1/search supporting keyword sorting and semantic vector thresholds.

Comparison: SkillNet vs. Ad-hoc Prompts vs. Static Registries

CLI Quickstart & Python SDK

Install the Python package and launch the local exploration UI:

# Install SkillNet with local UI support
pip install "skillnet-ai[ui]"

# Search the public skill library
skillnet search "analyze financial PDF reports" --limit 5

# Launch the offline local browser interface
skillnet ui --skills-dir ~/.claude/skills

Query the SkillNet catalog programmatically via Python:

from skillnet_ai import SkillNetClient

client = SkillNetClient()

# Semantic vector search across 500K+ skills
results = client.search(
    q="extract structured tables from invoice images",
    mode="vector",
    threshold=0.85,
    limit=5
)

for skill in results:
    print(f"Skill: {skill.skill_name} | Stars: {skill.stars} | URL: {skill.skill_url}")

Claude Code & MCP Integration

Register the SkillNet MCP server to enable autonomous skill discovery inside Claude Code or Cursor:

# Clone and install the SkillNet MCP server
git clone https://github.com/CycleChain/skillnet-mcp
cd skillnet-mcp
npm install

Add to your .mcp.json:

{
  "mcpServers": {
    "skillnet": {
      "command": "node",
      "args": ["/path/to/skillnet-mcp/build/index.js"],
      "env": {}
    }
  }
}

Technical Specifications

Frequently Asked Questions