DuckDuckGo Search MCP Review 2026: Privacy-First Search for AI Agents

By Wayne MacDonald - April 17, 2026 - 11 min read - Reviews

What is the DuckDuckGo Search MCP?

The DuckDuckGo & Felo AI Search MCP is an open-source Model Context Protocol server maintained by OEvortex. It gives any MCP-compatible AI agent - Claude Desktop, Cline, Cursor, Windsurf - the ability to search the web privately, fetch page content, and get AI-synthesized answers, all without needing an API key or account.

If you've been burned by API-key-gated search plugins (looking at you, Exa and Tavily free tiers), this is the plugin you reach for when you just want your agent to google something and move on.

Two things make it stand out in April 2026:

1. Felo AI bundling. Most search MCPs return a list of links. This one optionally hands you a synthesized, cited answer via Felo - closer to what Perplexity does, but free and MCP-native.
2. No-key philosophy. The server scrapes DuckDuckGo's public HTML with rotating user agents and rate limiting. You don't manage a secret; your agent doesn't leak one.

Six core features

1. DuckDuckGo web search (web-search)

Scrapes DuckDuckGo's HTML results page. Returns title, URL, and snippet for each hit. Supports filtering by number of results and region. Under light load (50 queries/hour), I saw sub-700ms latency from a US East container.

2. Felo AI search (felo-search)

Pipes your query to Felo's answer engine and returns a synthesized response with inline citations. Supports both standard (wait-then-return) and streaming modes. I used it to ask "what's new in Next.js 15.5 SSG" and got a paragraph with five cited source URLs - exactly what I'd want an agent to receive.

3. URL content extraction (fetch-url-content)

Give it a URL, get clean article text back. Supports maxLength caps and excludeTags (nav, footer, aside) so you're not paying token bills for boilerplate. This is the tool that turns search + read workflows into one-hop MCP calls.

4. URL metadata extraction (fetch-url-metadata)

Returns OpenGraph/Twitter card data: title, description, image, site name. Useful when an agent is building a citation list or summarizing a link roll.

5. Privacy-first architecture

No analytics, no logs, no phone-home. The server rotates user agents and implements polite rate limiting. DuckDuckGo itself doesn't store your queries - and since this runs locally via npx, neither does any middleman.

6. Smart caching and transports

In-memory LRU cache for repeat queries (configurable TTL) and support for stdio, SSE, and Streamable HTTP transports. That means you can self-host it as a long-lived HTTP service or embed it inline with Claude Desktop - same binary.

Install in 60 seconds

Add this to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on Mac, %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "ddg_search": {
      "command": "npx",
      "args": ["-y", "@oevortex/ddg_search@latest"]
    }
  }
}

Restart Claude. That's it. Test it with: "Search the web for the latest MCP spec updates and summarize." The agent will pick up web-search automatically.

For Cursor or Cline, the same JSON block goes into each tool's MCP settings panel. For Smithery one-click install:

npx -y @smithery/cli install @OEvortex/ddg_search --client claude

Pricing

Free. Forever. Released under Apache 2.0. No tiers, no usage caps enforced by the server itself, no premium add-ons. You can literally fork it and run your own private mirror if you want.

The only implicit "cost" is compute - it runs in your own Node process or Docker container - and politeness limits on how aggressively you scrape. For personal-agent use (even heavy use), you'll never hit a wall.

Pros and cons

Alternatives compared

• Brave Search MCP - free tier caps at 2K queries/month, needs API key. Better SLA, no URL fetch tool. Full Brave review.
• Exa MCP - neural semantic search, $10/1K queries, closed-source SaaS. Best when you need embedding-style retrieval, worst for cost-sensitive solo devs.
• Tavily MCP - RAG-optimized, 1K free/month then paid. Returns clean structured JSON for agent pipelines. Not self-hostable.
• Google Custom Search MCP - requires CSE ID and API key, 100 free queries/day. Great authority, brutal quota, painful setup.

If you only install one free search MCP in 2026, it's this one. If you need semantic/neural search or enterprise guarantees, pair it with Exa or Tavily.

FAQ

Verdict: 4.3 / 5

The OEvortex DuckDuckGo + Felo AI Search MCP is the first search plugin I install on any new machine. It's free, it's private, it's MCP-native, and it bundles AI synthesis most competitors charge for. The half-star off is realism: scraping is inherently fragile, and the maintainers can't promise an SLA. For anything mission-critical, layer it with Brave or Tavily as a backup.

But for 90% of "agent needs to look something up on the web" jobs? Start here.

Reviewed by Wayne MacDonald, Skiln.co MCP editor. Independent review - no affiliate or sponsorship from OEvortex or DuckDuckGo.