web-search-mcp
Provides multi-provider web search capabilities with fallback chains, semantic reranking, and content extraction for grounded agent retrieval.
README
web-search-mcp
Best-in-class web search MCP server for grounded agent retrieval.
Features
- Multi-provider search (
tavily,brave,exa,serper,searxng,duckduckgo) - Provider fallback chain and optional ensembling
- Optional local semantic rerank (
RERANKER=local) with@xenova/transformers - SSRF-safe URL fetching with robots policy support
- Injection defense:
<untrusted_content>wrapping, hidden-char stripping, heuristic detection - HTML, PDF, RSS, and YouTube transcript extraction
- Reranking, dedupe/canonicalization, chunk anchors for citation
- Persistent search/page memory with SQLite (
STATE_BACKEND=sqlite) - MCP-native capabilities: tools, resources, prompts, completions, sampling, elicitation
- Evaluation harness (
Recall@5,MRR@10) and provider A/B compare CLI - Production endpoints:
/livez,/readyz,/metrics,/health
Install
npm i
npm run build
Configure via .env
Copy and fill the template:
cp .env.example .env
Supported keys:
SEARCH_PROVIDER=tavily|brave|exa|serper|duckduckgoTAVILY_API_KEY,BRAVE_API_KEY,EXA_API_KEY,SERPER_API_KEYSEARCH_PROVIDER_FALLBACK=brave,duckduckgoHOST=127.0.0.1STATE_BACKEND=memory|sqlite|redisSQLITE_PATH=.cache/web-search-mcp.sqliteDOMAIN_ALLOWLIST=*.gov,*.eduDOMAIN_BLOCKLIST=internal.example.comHTTP_RATE_LIMIT_RPS=30FETCH_MAX_BODY_BYTES=5000000PORT=8080MCP_BASE_PATH=/mcp(set/mcp/searchif desired)CORS_ORIGIN=*(or a specific origin)MCP_AUTH_TOKEN=<secret>(optional bearer token for MCP endpoint)DEBUG=0|1
Run locally (default streamable MCP)
npm run dev
This auto-loads .env and starts a streamable MCP endpoint at:
http://127.0.0.1:8080/mcp(orHOST/PORT+MCP_BASE_PATHfrom.env)
If you need stdio mode explicitly:
npm run dev:stdio
MCP client configuration
Streamable MCP (recommended)
Use URL-based config:
{
"mcpServers": {
"web-search": {
"url": "http://127.0.0.1:8080/mcp"
}
}
}
If you set MCP_BASE_PATH=/mcp/search, use:
{
"mcpServers": {
"web-search": {
"url": "http://127.0.0.1:8080/mcp/search"
}
}
}
If MCP_AUTH_TOKEN is enabled, include:
Authorization: Bearer <token>
Stdio fallback (.cursor/mcp.json)
If your client requires command/stdio mode, use:
{
"mcpServers": {
"web-search": {
"command": "npx",
"args": ["-y", "web-search-mcp", "--stdio"],
"env": {
"SEARCH_PROVIDER": "tavily",
"TAVILY_API_KEY": "YOUR_KEY",
"SEARCH_PROVIDER_FALLBACK": "brave,duckduckgo"
}
}
}
}
Claude Desktop (stdio fallback)
{
"mcpServers": {
"web-search": {
"command": "npx",
"args": ["-y", "web-search-mcp", "--stdio"],
"env": {
"SEARCH_PROVIDER": "tavily",
"TAVILY_API_KEY": "YOUR_KEY"
}
}
}
}
VS Code MCP (stdio fallback)
{
"servers": {
"web-search": {
"command": "npx",
"args": ["-y", "web-search-mcp", "--stdio"],
"env": {
"SEARCH_PROVIDER": "duckduckgo"
}
}
}
}
Tools
searchfetch_urlsearch_and_extractnews_searchacademic_searchget_youtube_transcriptextract_pdfcheck_config
Prompts
research_topicfact_checklatest_news
Resources
web-search://page/{id}web-search://history/{id}
CLI
npm run compare -- --query "model context protocol"
npm run eval
Inspector validation
npx @modelcontextprotocol/inspector npm run dev
Smoke test
curl -s http://127.0.0.1:8080/health
curl -s http://127.0.0.1:8080/livez
curl -s http://127.0.0.1:8080/readyz
curl -s http://127.0.0.1:8080/metrics
Expected:
{ "status": "ok", "version": "0.1.0", "provider": "..." }
With auth enabled:
curl -i http://127.0.0.1:8080/mcp
# 401 unauthorized (without Bearer token)
Docker
Build:
docker build -t web-search-mcp .
Run with env file:
docker run --env-file .env -p 8080:8080 web-search-mcp
Then validate each tool:
searchfetch_urlsearch_and_extractnews_searchacademic_searchget_youtube_transcriptextract_pdfcheck_config
For nginx/Kubernetes deployment guidance see docs/NGINX.md and manifests in deploy/k8s plus Helm chart in charts/web-search-mcp.
Build and release
npm test
npm run build
Publish:
npm publish --provenance
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。