Firecrawl

Firecrawl’s first-party MCP server hands an agent search, scrape, and page-interaction tools that return content already cleaned up for LLM consumption — markdown instead of raw HTML, with boilerplate and navigation chrome stripped out. That matters more than it sounds: an agent doing multi-step research spends a meaningful share of its context budget parsing pages, and pre-cleaned output cuts that overhead directly.

Configuration is a single required variable, FIRECRAWL_API_KEY, against Firecrawl’s cloud service; FIRECRAWL_API_URL swaps that for a self-hosted deployment instead. A handful of secondary variables control transport mode (HTTP_STREAMABLE_SERVER for local streaming instead of stdio) and surface area (FIRECRAWL_MCP_SEARCH_ENABLED to restrict the server to search-only), which is useful when an integration only needs one capability and shouldn’t have the rest available.

It’s documented against Claude Desktop, Cursor 0.45.6+, Windsurf, and VS Code / VS Code Insiders, plus the MCP.so playground for quick testing — broad enough coverage that it’s a reasonable default whenever a task calls for live web content rather than a static, pre-indexed search API.