DuckDuckGo MCP Server

MCP server that gives Claude and other LLMs DuckDuckGo web search plus URL content fetching, with no API key required.

Free· Free and open source (MIT). DuckDuckGo's search endpoint is used without API keys, so there are no per-query costs; self-hosting only incurs whatever compute you run it on.MCP Servers

Visit website →

Best for

Developers and hobbyists who want to give Claude Desktop or a custom MCP agent free, no-account web search and lightweight URL reading without wiring up a paid SERP API.

Skip if

Production agents doing high-volume research, or workflows that need Google-grade result quality, structured SERP metadata, or guaranteed uptime and SLAs.

DuckDuckGo MCP Server is a lightweight, MIT-licensed Model Context Protocol server written in Python that plugs web search and page-scraping capabilities into any MCP-aware client — Claude Desktop, Claude Code, Cursor, Continue, or a custom agent using an MCP SDK. It exposes two tools: a search tool that hits DuckDuckGo's public HTML endpoint with configurable result count, region, and SafeSearch level, and a fetch tool that pulls a URL and returns cleaned, paginated text extracted from the page's HTML so the model can reason over content that was not in its training data.

The project is one of the most popular community MCP servers (around 1.4k GitHub stars) largely because it removes the two biggest friction points of giving an LLM live search: no Bing/Serper/SerpAPI account to provision, and no per-query cost. It ships with sensible defaults — 30 searches/minute and 20 fetches/minute of built-in rate limiting to stay under DuckDuckGo's soft blocks, SSRF protection that refuses private/loopback targets, and an optional `[browser]` extra that swaps in a headless browser backend when DDG's fingerprint filter starts returning empty results. Installation is a one-liner with `uv pip install duckduckgo-mcp-server` plus a few JSON lines in your client's MCP config; SSE, streamable-HTTP, and Docker transports are also documented for remote deployments.

Typical workflows: giving Claude Desktop the ability to answer questions about events after its knowledge cutoff, letting a coding agent look up current library docs before writing code, powering a research subagent that searches then fetches the top N results, or acting as the search leg of a home-grown deep-research loop where a planner model iterates queries against the tool.

Editor's take

This is the MCP server I recommend first when someone asks how to get Claude on the live web without a bill. It is not the strongest search backend you can wire up — a paid Serper or Tavily connector will beat it on quality and reliability — but the ergonomics of 'uv pip install, paste four lines into your config, done' are hard to argue with, and the built-in rate limits mean beginners are unlikely to get themselves blacklisted.

— The AI Tool Bible editorial team

Pros

✅ Zero configuration and zero cost — no API keys, accounts, or billing setup
✅ MIT-licensed and self-hostable, so nothing about your query stream leaves your machine except the DuckDuckGo request itself
✅ Two clean tools (search + fetch) that map onto the way agents actually use the web — search, then read the interesting hit
✅ Built-in per-minute rate limits and SSRF guardrails ship enabled by default
✅ Works out of the box with Claude Desktop, Claude Code, and any other MCP client via stdio, SSE, or streamable HTTP
✅ Optional browser backend for when DuckDuckGo's bot detection starts serving empty pages
✅ Small, focused Python codebase that is easy to fork and modify

Cons

⚠️ DuckDuckGo's public endpoint is rate-limited and fingerprint-filtered — heavy or bursty use will get you empty results or temporary blocks even with the browser backend
⚠️ Search quality is DDG-quality: solid for general queries, weaker than Google/Bing for long-tail technical or academic lookups
⚠️ The fetch tool extracts plain text and drops most structure, so it is not a substitute for a real scraping/RAG pipeline on complex pages
⚠️ No caching or deduplication layer — every call re-hits DDG, which matters when an agent loops
⚠️ Community-maintained by a single developer; support and roadmap depend on the maintainer's availability

Use cases

Adding live web search to Claude DesktopGrounding an MCP-based research agentFetching and summarising a specific URL from chatLocal deep-research loops without a SERP API billCoding agents looking up current library documentationFact-checking questions past the model's knowledge cutoffPrototyping RAG pipelines before paying for a search API