SearXNG HTTP MCP

中文 · Quick Start · Features · Architecture · Comparison · Usage · MCP Tools · Client Config · Plugin · Contributing

A self-contained MCP server that wraps SearXNG — a free, privacy-respecting metasearch engine that aggregates results from 200+ search engines.

---

🚀 Quick Start

Server mode — deploy once, connect from any client:

docker run -d --name searxng-mcp --restart unless-stopped \

-p YOUR_PORT:8888 --memory=512m --cpus=1 \ ghcr.io/whw23/searxng-http-mcp:latest

Then connect your client to http://YOUR_HOST:YOUR_PORT/mcp/. To enable API key auth, see Authentication.

Local mode — no server needed, run directly in your client:

docker run --rm -i --memory=512m --cpus=1 ghcr.io/whw23/searxng-http-mcp:latest --stdio

Add this as a stdio MCP server in your client — see Client Configuration for details.

uvx mode — if you already have SearXNG running (install guide):

uvx searxng-http-mcp

Set SEARXNG_URL to point to your SearXNG instance (default: http://127.0.0.1:8080).

✨ Features

Search

• 🔍 200+ search engines — Google, Bing, DuckDuckGo, Brave, and more via SearXNG
• 📂 30+ categories — news, images, videos, science, IT, and more
• 📄 Multi-page fanout — up to 5 pages per call
• 💡 Autocomplete suggestions — discover relevant search terms
• 🗂 Engine discovery — query available engines grouped by category
• 🎯 Token-efficient — results trimmed to essentials

Infrastructure

• 📦 Self-contained — SearXNG built into Docker image
• 🔄 Triple transport — HTTP server, Docker stdio, and uvx standalone
• 🔐 Authentication — x-api-key + HTTP Basic Auth
• 🌐 Reverse proxy — SearXNG Web UI on the same port
• ⚡ Dynamic tool descriptions — live category lists injected at startup
• 📐 Rich JSON Schema — enum constraints, range limits, and descriptions on every parameter

🏛 Architecture

graph LR

Client(["client:YOUR_PORT"]) --> Expose(":8888")

subgraph Container["🐳 Docker Container"] direction LR Expose --> Auth{Auth} Auth -->|/mcp| MCP[FastMCP Server] Auth -->|/*| Proxy[Reverse Proxy] MCP --> SearXNG[SearXNG :8080] Proxy --> SearXNG end

style Expose fill:none,stroke:#2496ed,stroke-dasharray:5 5,color:#2496ed

style Client fill:#4a90d9,color:#fff,stroke:#3a7bc8 style Container fill:#f0f4f8,stroke:#2496ed,stroke-width:2px,color:#2496ed style Auth fill:#f5a623,color:#fff,stroke:#d4900e style MCP fill:#50c878,color:#fff,stroke:#3da85e style Proxy fill:#9b59b6,color:#fff,stroke:#8344a5 style SearXNG fill:#e74c3c,color:#fff,stroke:#c0392b

📊 Comparison with Alternatives

Why these five?

There are 20+ SearXNG MCP servers and many more general-purpose search MCPs. Most SearXNG wrappers only expose a basic search tool, leaving SearXNG's categories, autocomplete, and engine metadata unused. We picked five alternatives that each represent a distinct category:

• 88plug/searxng-mcp — richest tool surface among SearXNG MCPs (7 tools: rendered fetch, research mode, parallel queries)
• ihor/mcp-searxng — most GitHub stars among SearXNG MCPs
• open-webSearch — top free multi-engine alternative outside the SearXNG ecosystem (Bing, Baidu, DuckDuckGo, Brave, etc.)
• exa-mcp-server — most popular commercial search API MCP
• Perplexity MCP — commercial AI-powered search, highest star count in the search MCP space

Feature ✨ This project 88plug/searxng-mcp ihor/mcp-searxng open-webSearch exa-mcp-server Perplexity MCP

Search 200+ engines via SearXNG ✅ ✅ ✅ ❌ ❌ ❌ 30+ search categories ✅ ❌ ❌ ❌ ❌ ❌ Multi-page fanout ✅ ❌ ❌ ❌ ❌ ❌ Autocomplete suggestions ✅ ❌ ❌ ❌ ❌ ❌ Engine discovery tool ✅ ❌ ❌ ❌ ❌ ❌ Dynamic tool descriptions ✅ ❌ ❌ ❌ ❌ ❌ Infrastructure Self-contained (built-in search) ✅ ❌ ❌ ✅ N/A N/A Zero-install Docker deploy ✅ ❌ ❌ ✅ ❌ ❌ HTTP + stdio transport ✅ ✅ ✅ ✅ ✅ ✅ Authentication ✅ ❌ ❌ ❌ ✅ ✅ Web UI reverse proxy ✅ ❌ ❌ ❌ ❌ ❌ AI Coding Agent Plugin ✅ ❌ ❌ ❌ ❌ ✅ General Free & open source ✅ ✅ ✅ ✅ ❌ (paid API) ❌ (paid API) Privacy (self-hosted) ✅ ✅ ✅ ✅ ❌ ❌ Language Python Python Node.js TypeScript TypeScript TypeScript GitHub Stars

Why fewer tools?

MCP is designed for composition — clients connect multiple specialized servers, each doing one thing well. Some alternatives bundle URL fetching, rendered page extraction, multi-query fan-out, or research modes into the search server. We k