One Search MCP Server

Web search in OneSearch starts with one_search, which returns result URLs, titles, and descriptions from the configured provider. The default uses local browser automation; alternatives include DuckDuckGo, SearXNG, and Tavily. Local search and DuckDuckGo need no API key, while Tavily requires one. For link discovery, one_map loads a supplied page in the browser and collects links from its HTML rather than crawling a sitemap.

Content capture and preparation use separate tools. one_scrape captures an individual page as Markdown, HTML, links, or screenshots, with a full-page screenshot option. Before capture, it can run waits, clicks, text entry, key presses, and scrolling in sequence, stopping at the first action error. JavaScript actions require allowExecuteJavascript to be true. one_extract prepares cleaned text from supplied URLs for other tools or models instead of performing LLM extraction. Browser operations run locally through agent-browser and need a Chromium-based browser; the Docker image includes Chromium. The project provides client configuration examples for Claude Code, Claude Desktop, Cursor, and Windsurf.