🔎 @caikiji/mcp-search

MCP server for web search — privacy-respecting, self-hosted metasearch

Installation · Configuration · Tools · 中文文档

✨ Features

• 🌐 Web search — aggregates results from multiple engines (Brave, DuckDuckGo, Bing News, etc.)
• 📄 Page fetch — read full page content from search results as Markdown
• 🔍 Engine discovery — info to see engine counts; info scope: "engines" to list all
• 🎯 Token efficient — compact mode, smart snippet truncation, deduplicated results
• 🔒 Self-hosted — works with your own SearXNG instance, full privacy control

Installation

npm install -g @caikiji/mcp-search

Configuration

Environment variables

Variable	Default	Description
SEARCH_URL	—	Your SearXNG instance URL (required)
SEARCH_AUTH	—	Basic auth credentials (user:pass). Leave empty if your instance is public.
SEARCH_TIMEOUT	15000	SearXNG API timeout (ms)
SEARCH_FETCH_TIMEOUT	15000	Page fetch timeout (ms)
SEARCH_DEFAULT_COUNT	10	Default results per search (1–50)
SEARCH_MAX_LENGTH	8000	Max chars for result content (500–50000)
SEARCH_SNIPPET_LENGTH	300	Max chars for result snippet (truncated at word boundary)
SEARCH_FETCH_UA	Chrome UA	User-Agent for page fetching
SEARCH_RECOMMENDED_ENGINES	—	Semicolon-separated recommended engines, optionally with engine|description. Shown in info output. See config example below.

MCP Client config

{
  "mcpServers": {
    "search": {
      "command": "npx",
      "args": ["-y", "@caikiji/mcp-search"],
      "env": {
        "SEARCH_URL": "https://search.example.com",
        "SEARCH_AUTH": "user:password",
        "SEARCH_RECOMMENDED_ENGINES": "duckduckgo|general purpose, privacy;bing|images, news;sogou|Chinese web;sogou wechat|WeChat articles;stackoverflow|programming Q&A;github|code repositories"
      }
    }
  }
}

Tools

Search

Tool	Arguments	Description
query	query, [language], [categories], [time_range], [engines], [pageno], [count], [format]	Web search with dedup, snippet, and source info. format: "compact" for minimal token usage.
result	url, [max_length]	Fetch a URL from results and return as Markdown. Works on most blogs/docs sites.
info	[scope]	Instance info. Default: counts, categories, settings. scope: "engines" for full list with categories.

query parameters

Param	Description
query	Search query (required)
site	Restrict to domain, e.g. github.com. Prepends site: to the query.
language	Language code (zh-CN, en-US, auto). Default: auto
categories	Comma-separated: general, news, images, video, music, it, science, files, social media
time_range	day, week, month, year
engines	Comma-separated engine names. Use info to see available ones
pageno	Page number. Default: 1
count	Results to return (1–50). Default: 10
format	full (title+URL+snippet) or compact (title+URL only, minimal tokens). Default: full
depth	When true, forces count=1 and fetches the top result's full content via reader mode. Saves the query→pick→fetch round trip. Default: false

Output format

Full mode:

5 results | Sources: duckduckgo,brave
[1] Title
URL: https://...
Src: duckduckgo

Snippet text here (truncated to 300 chars)...

[2] Next Title
URL: https://...
Src: google +2

Compact mode:

5 results | Sources: duckduckgo
[1] Title — https://...
[2] Next Title — https://...

---

📖 中文文档