skill-seekers-reference/docs/zh-CN/advanced/mcp-server.md

MCP Server Setup Guide

Skill Seekers v3.2.0
通过 Model Context Protocol 与 AI 代理集成

---

What is MCP?

MCP (Model Context Protocol) lets AI agents like Claude Code control Skill Seekers through natural language:

You: "Scrape the React documentation"
Claude: ▶️ scrape_docs({"url": "https://react.dev/"})
        ✅ Done! Created output/react/

---

Installation

# Install with MCP support
pip install skill-seekers[mcp]

# Verify
skill-seekers-mcp --version

---

Transport Modes

stdio Mode (Default)

For Claude Code, VS Code + Cline:

skill-seekers-mcp

Use when:

• Running in Claude Code
• Direct integration with terminal-based agents
• Simple local setup

---

HTTP Mode

For Cursor, Windsurf, HTTP clients:

# Start HTTP server
skill-seekers-mcp --transport http --port 8765

# Custom host
skill-seekers-mcp --transport http --host 0.0.0.0 --port 8765

Use when:

• IDE integration (Cursor, Windsurf)
• Remote access needed
• Multiple clients

---

Claude Code Integration

Automatic Setup

# In Claude Code, run:
/claude add-mcp-server skill-seekers

Or manually add to ~/.claude/mcp.json:

{
  "mcpServers": {
    "skill-seekers": {
      "command": "skill-seekers-mcp",
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "GITHUB_TOKEN": "ghp_..."
      }
    }
  }
}

Usage

Once connected, ask Claude:

"List available configs"
"Scrape the Django documentation"
"Package output/react for Gemini"
"Enhance output/my-skill with security-focus workflow"

---

Cursor IDE Integration

Setup

1. Start MCP server:

skill-seekers-mcp --transport http --port 8765

2. In Cursor Settings → MCP:
   • Name: skill-seekers
   • URL: http://localhost:8765

Usage

In Cursor chat:

"Create a skill from the current project"
"Analyze this codebase and generate a cursorrules file"

---

Windsurf Integration

Setup

1. Start MCP server:

skill-seekers-mcp --transport http --port 8765

2. In Windsurf Settings:
   • Add MCP server endpoint: http://localhost:8765

---

可用工具

27 个工具，按类别组织：

核心工具（9 个）

• list_configs - 列出预设
• generate_config - 从 URL 创建配置
• validate_config - 检查配置
• estimate_pages - 页面估算
• scrape_docs - 抓取文档
• package_skill - 打包技能
• upload_skill - 上传到平台
• enhance_skill - AI 增强
• install_skill - 完整工作流

扩展工具（10 个）

• scrape_github - GitHub 仓库
• scrape_pdf - PDF 提取
• scrape_generic - 10 种新来源类型的通用抓取器（见下文）
• scrape_codebase - 本地代码
• unified_scrape - 多源抓取
• detect_patterns - 模式检测
• extract_test_examples - 测试示例
• build_how_to_guides - 操作指南
• extract_config_patterns - 配置模式
• detect_conflicts - 文档/代码冲突

配置源（5 个）

• add_config_source - 注册 Git 源
• list_config_sources - 列出源
• remove_config_source - 删除源
• fetch_config - 获取配置
• submit_config - 提交配置

向量数据库（4 个）

• export_to_weaviate
• export_to_chroma
• export_to_faiss
• export_to_qdrant

scrape_generic 工具

scrape_generic 是 v3.2.0 新增的 10 种来源类型的通用入口。它将请求委托给相应的 CLI 抓取器模块。

支持的来源类型： jupyter（Jupyter 笔记本）、html（本地 HTML）、openapi（OpenAPI/Swagger 规范）、asciidoc（AsciiDoc 文档）、pptx（PowerPoint 演示文稿）、rss（RSS/Atom 订阅源）、manpage（Man 手册页）、confluence（Confluence 维基）、notion（Notion 页面）、chat（Slack/Discord 聊天记录）

参数：

名称	类型	必需	描述
source_type	string	是	10 种支持的来源类型之一
name	string	是	输出的技能名称
path	string	否	文件或目录路径（用于基于文件的来源）
url	string	否	URL（用于 confluence、notion、rss 等基于 URL 的来源）

使用示例：

"抓取 Jupyter 笔记本 analysis.ipynb"
→ scrape_generic(source_type="jupyter", name="analysis", path="analysis.ipynb")

"提取 API 规范内容"
→ scrape_generic(source_type="openapi", name="my-api", path="api-spec.yaml")

"处理 PowerPoint 演示文稿"
→ scrape_generic(source_type="pptx", name="slides", path="presentation.pptx")

"抓取 Confluence 维基"
→ scrape_generic(source_type="confluence", name="wiki", url="https://wiki.example.com")

详见 MCP 参考文档。

---

Common Workflows

Workflow 1: Documentation Skill

User: "Create a skill from React docs"
Claude: ▶️ scrape_docs({"url": "https://react.dev/"})
        ⏳ Scraping...
        ✅ Created output/react/
        
        ▶️ package_skill({"skill_directory": "output/react/", "target": "claude"})
        ✅ Created output/react-claude.zip
        
        Skill ready! Upload to Claude?

Workflow 2: GitHub Analysis

User: "Analyze the facebook/react repo"
Claude: ▶️ scrape_github({"repo": "facebook/react"})
        ⏳ Analyzing...
        ✅ Created output/react/
        
        ▶️ enhance_skill({"skill_directory": "output/react/", "workflow": "architecture-comprehensive"})
        ✅ Enhanced with architecture analysis

Workflow 3: Multi-Platform Export

User: "Create Django skill for all platforms"
Claude: ▶️ scrape_docs({"config": "django"})
        ✅ Created output/django/
        
        ▶️ package_skill({"skill_directory": "output/django/", "target": "claude"})
        ▶️ package_skill({"skill_directory": "output/django/", "target": "gemini"})
        ▶️ package_skill({"skill_directory": "output/django/", "target": "openai"})
        ✅ Created packages for all platforms

---

Configuration

Environment Variables

Set in ~/.claude/mcp.json or before starting server:

export ANTHROPIC_API_KEY=sk-ant-...
export GOOGLE_API_KEY=AIza...
export OPENAI_API_KEY=sk-...
export GITHUB_TOKEN=ghp_...

Server Options

# Debug mode
skill-seekers-mcp --verbose

# Custom port
skill-seekers-mcp --port 8080

# Allow all origins (CORS)
skill-seekers-mcp --cors

---

Security

Local Only (stdio)

# Only accessible by local Claude Code
skill-seekers-mcp

HTTP with Auth

# Use reverse proxy with auth
# nginx, traefik, etc.

API Key Protection

# Don't hardcode keys
# Use environment variables
# Or secret management

---

Troubleshooting

"Server not found"

# Check if running
curl http://localhost:8765/health

# Restart
skill-seekers-mcp --transport http --port 8765

"Tool not available"

# Check version
skill-seekers-mcp --version

# Update
pip install --upgrade skill-seekers[mcp]

"Connection refused"

# Check port
lsof -i :8765

# Use different port
skill-seekers-mcp --port 8766

---

See Also

• MCP 参考文档 - 完整工具参考
• MCP 工具深入 - 高级用法
• MCP 协议 - 官方 MCP 文档