DuckDuckGo MCP Server

Executes web searches against DuckDuckGo's HTML interface (not API-based) and returns formatted results with titles, URLs, and snippets optimized for LLM consumption. The implementation queries DuckDuckGo directly without requiring API keys, removes ad content and cleans redirect URLs before returning results. Results are rate-limited to 30 requests per minute to prevent service abuse.

Fetches raw HTML from a specified URL and parses it into cleaned, LLM-consumable text content. The implementation uses HTTP requests to retrieve webpages, applies HTML parsing to extract meaningful content while removing boilerplate (scripts, styles, navigation), and formats the output as plain text. Rate-limited to 20 requests per minute to prevent overloading target servers.

Implements per-tool rate limiting using a quota system: 30 requests per minute for search, 20 requests per minute for content fetching. The implementation tracks request timestamps and enforces limits before executing tool methods, returning rate-limit errors when quotas are exceeded. This prevents both external service abuse and protects against runaway LLM agent loops.

Exposes search and content-fetching capabilities as MCP tools using the FastMCP framework, which handles tool schema generation, parameter validation, and client communication. Tools are registered via @mcp.tool() decorators that automatically generate JSON schemas for parameters (query, max_results, url) and integrate with any MCP-compatible client. The server runs as a standalone process that clients connect to via stdio or network transport.

Implements comprehensive error catching and reporting for network failures, malformed URLs, unreachable servers, and parsing errors. When requests fail (timeout, connection error, 404, etc.), the system returns descriptive error messages to the LLM client rather than crashing. This allows LLM agents to handle failures programmatically (retry, try alternative queries, etc.) rather than terminating.

Allows clients to specify the maximum number of search results to return via the max_results parameter (default: 10). The implementation respects this parameter when querying DuckDuckGo and truncates results before formatting and returning them. This enables clients to balance between result comprehensiveness and token consumption in LLM prompts.