The Firecrawl MCP Server is an official Model Context Protocol server that integrates Firecrawl's web scraping and search capabilities into LLM clients including Cursor, Claude, and other MCP-compatible AI agents. Written in JavaScript, the project enables AI systems to search, scrape, and interact with live web content to gather clean, structured data for agent-ready context.
The repository provides multiple deployment options to accommodate different user needs. Users can connect to a hosted keyless free tier without setup, run the server locally via npx, or perform manual installation. The project includes specific configuration guides for popular IDEs and editors, with detailed instructions for Cursor versions 0.45.6 and above, Windsurf, VS Code, and Claude Desktop. For developers preferring local HTTP streaming instead of the default stdio transport, the server supports Streamable HTTP Local Mode on port 3000. The project also maintains legacy support through Smithery for one-click installation to Claude Desktop.
The core functionality centers on eight primary tools accessible through the MCP interface. The scrape tool extracts content from single URLs with support for JSON format with custom schemas, markdown format for full page content, and a specialized branding format that captures comprehensive brand identity including colors, fonts, typography, and UI components. The map tool discovers indexed URLs on websites, the search tool queries the web and extracts content from results, and the interact tool enables browser automation including clicking, typing, and navigation. Additional tools include crawl for multi-page extraction with configurable limits, parse for processing files and hosted uploads, extract for structured data extraction, agent for complex multi-source research, monitor for recurring page checks, and research for paper and GitHub repository analysis.
Configuration flexibility is a key design feature. The server supports cloud API usage via Firecrawl API keys and self-hosted instances through custom API endpoints. It implements OAuth bearer token authentication for hosted Firecrawl, accepting access tokens prefixed with fco_ while rejecting refresh tokens. Environment variables control deployment mode, with FIRECRAWL_API_KEY for cloud API, FIRECRAWL_API_URL for self-hosted instances, and FIRECRAWL_OAUTH_TOKEN for static OAuth access tokens.
According to GitGenius activity tracking across 76 issues and pull requests, the repository maintains a median response latency of 15.3 hours with a mean of 383.2 hours, indicating active but variable engagement patterns. The most active contributors are nickscamara with 22 tracked events, vrknetha with 15 events, and dareima with 4 events. The README credits vrknetha and knacklabs for the initial implementation. The project shares contributors with microsoft/vscode, meshery/meshery, and zitadel/zitadel repositories, suggesting cross-project collaboration within the broader developer ecosystem.
The documentation emphasizes practical tool selection guidance, providing a quick reference table and detailed decision trees for choosing between scrape, map, search, crawl, and agent tools based on specific use cases. The project distinguishes between single-page extraction, multi-page discovery, web search, interactive automation, and complex research scenarios. It explicitly warns against common mistakes such as passing multiple URLs to a single scrape call and recommends JSON format with schemas over markdown format to prevent context window overflow in LLM applications.
