Firecrawl MCP server in Cursor let your editor read the web.
Cursor's agent gets scrape, search, crawl, and extract tools from one mcp.json entry. The config takes a minute; the toggle and the green dot are where people get stuck.
Pulling live docs, checking a competitor's changelog, extracting structured data from a page your code consumes — Cursor's agent can do all of it once Firecrawl MCP Server is connected. The server (github.com/firecrawl/firecrawl-mcp-server, 6.5k stars) wraps Firecrawl's scraping API as MCP tools.
Cursor reads MCP servers from mcp.json — globally at ~/.cursor/mcp.json, or per project at .cursor/mcp.json. Same format either way; the project file wins for that workspace.
Get the API key
Sign up at firecrawl.dev and create an API key (prefix fc-). The npm package firecrawl-mcp is free and open source; the key authenticates the scraping calls it makes to Firecrawl's hosted service, and your usage is metered there.
Edit mcp.json
Either go through the UI — Cursor Settings, then MCP, then Add new global MCP server, which opens the file for you — or create it directly. Use ~/.cursor/mcp.json to make Firecrawl available everywhere, or .cursor/mcp.json inside a repo to scope it (and optionally commit it for your team).
{
"mcpServers": {
"firecrawl": {
"command": "npx",
"args": ["-y", "firecrawl-mcp"],
"env": {
"FIRECRAWL_API_KEY": "fc-YOUR-API-KEY"
}
}
}
}Enable the toggle and check the dot
Back in Settings > MCP you should now see firecrawl in the server list with an on/off toggle and a status dot. Flip the toggle on; a green dot means Cursor spawned the process and completed the MCP handshake, and the tool names appear under the entry. Yellow or red means the spawn failed — see below.
Test it in the agent: "scrape https://example.com/pricing and put the tiers in a table." Cursor will ask to run the tool unless you've allowed it. The server's expected tool list is published on its live listing at https://loomal.ai/marketplace/firecrawl-mcp-server.
One Cursor-specific consideration: the agent performs best with a focused tool surface, and Firecrawl registers several tools (single-page scrape, site map, crawl, search, extract). If you run many MCP servers at once, disable the ones you're not using in Settings > MCP so Firecrawl's tools stay within the set Cursor offers the model.
Troubleshooting in Cursor
Red or yellow dot: hover the entry for the error, then check the usual suspects. On macOS, Cursor launched from the dock may not inherit your shell PATH, so npx fails to resolve — relaunch from a terminal (`cursor .`) or set the absolute npx path as command. On Windows, wrap the command: `cmd /c npx -y firecrawl-mcp`.
Server listed but the agent ignores it: confirm the toggle is actually on (servers default to enabled when added, but a previous failure can leave them off), and that you're using the agent rather than plain inline chat. If you edited mcp.json by hand while Cursor was open, hit the refresh icon in the MCP settings or restart Cursor so it re-reads the file. Tools running but failing is almost always the fc- key — Cursor passes env exactly as written, so a pasted space breaks auth.
FAQ
How do I install Firecrawl in Cursor?
Create or edit ~/.cursor/mcp.json (global) or .cursor/mcp.json (per project) with a firecrawl entry under mcpServers — command npx, args ["-y", "firecrawl-mcp"], FIRECRAWL_API_KEY in env — then enable the toggle in Cursor Settings > MCP.
Should I use the global or project mcp.json?
Global (~/.cursor/mcp.json) if you want scraping in every workspace; project (.cursor/mcp.json) if only one repo needs it or you want teammates to get the server when they clone. The project file takes precedence inside that workspace.
What does the status dot in Cursor's MCP settings mean?
Green means the server process started and Cursor completed the MCP handshake — tools are usable. Yellow or red means the spawn or handshake failed; the most common causes are npx not being on Cursor's PATH and JSON syntax errors in mcp.json.
Does Cursor ask before every Firecrawl call?
By default the agent requests approval per tool call. You can allow specific tools to run without prompting, which makes multi-page crawls smoother — just keep in mind every scrape counts against your Firecrawl plan.
More MCP servers for Cursor.
Browse scraping, search, and dev-tool servers with live tool lists.