Short answer: An MCP (Model Context Protocol) server for web scraping is a small process that exposes scrape, crawl, map, and extract as standardised tools an AI client can call. Instead of writing custom HTTP glue, clients like Claude Desktop and Cursor discover the tools, see their JSON schemas, and invoke them as if they were local functions.
The Model Context Protocol is an open spec (originated by Anthropic, now multi-vendor) that lets an AI host talk to external tools over a typed JSON-RPC channel — either local stdio or remote HTTP/SSE. The host gets a list of available tools with input schemas; the model picks one and the host invokes it. From the model's point of view the tool is a function with named arguments and a structured return value.
A scraping-flavoured MCP server typically registers four to five tools, each backed by an underlying scraping API:
scrape — fetch a single URL as markdown or HTML.crawl — start a multi-page job and stream results back.map — enumerate URLs on a domain.extract — pull structured fields out of one or many URLs using an LLM.search — web search with optional inline scrape.The model never sees the HTTP wire format. It sees a tool called scrape with arguments {url, formats} and gets back markdown directly.
fastCRW ships crw-mcp as an npm package. Add it to a client config (Claude Desktop's claude_desktop_config.json, Cursor's MCP settings, or any other host) and the five tools above appear without further wiring. The server talks to either the hosted API at https://api.fastcrw.com or a local self-hosted instance at http://localhost:3000 via env var. See the MCP clients page for client-specific configuration snippets.