What is Firecrawl MCP? Web data for your assistant, setup, credits and pricing
You ask your assistant to search for something or read a page; Firecrawl MCP searches the web, reads the page, crawls the site and (with an account) clicks and types on pages. The actual work is done by Firecrawl's cloud API; the official MCP server is MIT-licensed and the core Firecrawl engine is AGPL-3.0. There is a keyless endpoint (search, page reading and file parsing only; limited daily per IP); all tools require an account sign-in or an API key. The free plan gives 1,000 credits a month (1 credit per page, 2 credits for a 10-result search) and needs no card. Web content it fetches is untrusted content.
- What it does
- Has your assistant search the web, read pages and crawl sites
- Type
- MCP server (remote connector; local npx or Docker also available)
- Where
- Claude Code, Codex, Cursor, VS Code, Windsurf, OpenCode, ChatGPT, Claude
- Key
- Account or API key for the full version
- Data
- Goes to US servers; pages are cached for 2 days
- Turkish
- Not stated in official sources
- Cost
- Limited without a key; Free 1,000 credits/month; Hobby $19/month
What does Firecrawl MCP do?
MCP is the standard that lets AI programs connect to outside tools. With Firecrawl MCP your assistant reads web pages itself: it fetches the content of a URL you know as clean text or as data in fields you define, searches the web without a URL, crawls a site's pages in bulk or extracts its URL list. With an account you can also have it click and type on a dynamic page, search ready-made indexes such as scientific papers, GitHub issues and US legislation, and monitor page changes.
Data goes to US servers; pages are cached for 2 days. Fetched content is untrusted content.
The documentation says 27 tools; the number of named tools is 28. The keyless endpoint offers only three: firecrawl_search, firecrawl_scrape and firecrawl_parse.
Who it is for: people doing web research, price tracking and content compilation. Scraping carries legal and ethical responsibility; see the "What does it access?" section below.
How do you set up Firecrawl MCP?
There are three routes: the remote endpoint hosted by Firecrawl, the npx package running on your computer, and Docker. The commands are taken verbatim from the official documentation. The remote URLs are for clients, not pages to open in a browser.
› claude mcp add --transport http firecrawl https://mcp.firecrawl.dev/v2/mcp-oauthThen open the /mcp command in Claude Code and select Firecrawl; sign in to Firecrawl in the browser and approve a team.
› claude mcp add --transport http firecrawl https://mcp.firecrawl.dev/v2/mcp› codex mcp add firecrawl --url https://mcp.firecrawl.dev/v2/mcp-oauthThen codex mcp login firecrawl. In Cursor and clients configured with JSON, add a firecrawl entry under mcpServers with the same address as url; VS Code uses .vscode/mcp.json and OpenCode uses opencode.json (remote type).
The documentation says ChatGPT and Claude use the built-in Firecrawl connector and gives "Add to ChatGPT" and "Add to Claude" links; the menu path and plan requirement are not in the documentation. The menu path on Claude's connector directory page is not clear in official sources.
› env FIRECRAWL_API_KEY=fc-YOUR-API-KEY npx -y firecrawl-mcpRequires Node.js 22 or later. You create the key at https://www.firecrawl.dev/app/api-keys; don't put it in the server URL or the chat, put it in the client's secret settings. For Firecrawl on your own server, use FIRECRAWL_API_URL; for local HTTP mode, HTTP_STREAMABLE_SERVER=true.
The documentation itself pins the version in its example (firecrawl-mcp@3.23.7); using a pinned version is possible.
Try the keyless endpoint first; if tools like crawl, map and interact don't show up, you need to connect with an account.
How do you use Firecrawl MCP?
Every call spends credits; for large jobs (site crawling) ask it to set a page limit first. If you don't give a limit on crawl, the default cap is 10,000 pages.
!Fetched web page content is untrusted content; hidden instructions on a page can mislead the assistant. There is a GitHub report (#481) that credit-spending tools have no approval prompt: keep tool approval on in your client.
Which programs does Firecrawl MCP work in?
| Claude CodeYes | With OAuth or the keyless endpoint. |
|---|---|
| Codex, Cursor, VS Code, Windsurf, OpenCodeYes | The documentation has a setup example for each. |
| Claude DesktopYes | An npx entry and FIRECRAWL_API_KEY in claude_desktop_config.json. |
| ChatGPT and Claude (built-in connector)Unclear | The links are in the documentation; the menu path and plan requirement are not stated. |
| Gemini CLIUnclear | The repository has an extension file but no setup instructions. |
What does Firecrawl MCP access, what data is sent, and is scraping legal?
Public content of web pages (scrape, crawl, map, search), search results and, with an account, credit usage. To parse a file you upload, the hosted server doesn't read your local file directly; it gives you a signed upload command that you run.
firecrawl_interact clicks, types and navigates on a live page and can save cookies and localStorage to a named profile. firecrawl_monitor creates recurring monitors in your account and sends a webhook or email on changes. firecrawl_agent runs automated multi-source research and spends credits.
firecrawl_monitor_delete permanently deletes a monitor in your account (documentation: should only be called if the user explicitly asks). It does not delete files on your own computer.
Queries, URLs, extraction prompts and uploaded files go to Firecrawl servers; according to the privacy policy the servers are in the US. Scraped pages are cached for 2 days by default (turned off with storeInCache:false). Zero data retention is only on the Enterprise plan. Even with local npx or Docker, the cloud API does the work, except with Firecrawl on your own server.
!The terms of service expect you to comply with sites' own rules; robots.txt behavior is described only in the crawl documentation ("ignoreRobotsTxt" only on Enterprise). Legal responsibility for collecting personal data is yours. GitHub Security Advisories is empty, OSV has no entries and the repository has no SECURITY.md; that does not mean "secure". There are open reports, and whether they were fixed is not clear in official sources: local file reading (#306), crawl URL validation (#210), a search tool marked "read-only" getting a page-click option (#311), no approval on credit-spending tools (#481).
Is Firecrawl MCP free? Plans and credits
The MCP server itself is free (MIT); the cost is your Firecrawl account's credits. The keyless trial is free but has daily per-IP request and credit limits (the numbers are not in the documentation; exceeding them returns 429). Prices are always in US dollars. The prices below were verified on the firecrawl.dev/pricing page.
- 1,000 credits a month (500 searches or 1,000 pages)
- No card required
- 2 concurrent requests
- Continuing after credits run out (HTTP 402)
- 5,000 credits a month (2,500 searches or 5,000 pages)
- 5 concurrent requests
- 100,000 credits a month
- 25 concurrent requests
- 500,000 and 1,000,000 credits a month
- Enterprise: custom, including zero data retention
| Operation | Credits (per Firecrawl documentation) |
|---|---|
| Page reading (scrape) | 1 credit/page; JSON extraction +4, PDF +1/page. |
| Search | 2 credits per 10 results; extra per page if the result pages are also read. |
| Site crawl (crawl) | 1 credit per page. |
| Site map (map) | Documentation is inconsistent: the billing page says 1 per call, the pricing page says 1 per page. |
| Page actions (interact) | Documentation is inconsistent: the billing page says 2 to 7 per minute, the pricing page says 2 per minute. |
Source: Firecrawl pricing page and billing documentation. The terms of service list "commercial purposes except as expressly permitted by Firecrawl" among prohibited uses; the documentation does not separately say whether a paid plan covers commercial use. This is not legal advice.
Common problems and fixes
| What happens | What to do |
|---|---|
| "spawn npx ENOENT" on Windows | Install Node.js 22+, find the npx.cmd path with where npx, point the client to that path and fully restart the client. |
| 401 on connection | If the URL is /v2/mcp-oauth, sign in again; if it is /v2/mcp, change the key or switch the URL to /v2/mcp-oauth and complete sign-in; then start a new session. |
| crawl, map or agent tools don't show up | The keyless endpoint offers only search, scrape and parse; connect with an account or key. A team policy may have disabled some optional tools. |
| 429 (rate limit) | Wait for the retry interval or switch from keyless to account-based use. |
| Reported by users firecrawl_search returns 400 when connected to Firecrawl on your own server | Open issue #471 (October 1, 2026): since 3.25.0 the client sends cloud-only fields; the user reported that 3.24.0 works. No official fix has been announced. |
How do you remove Firecrawl MCP?
- From the client:
claude mcp remove firecrawl, or delete thefirecrawlentry from the MCP settings file. - If you connected with an account: Revoke the connection at firecrawl.dev/app/settings?tab=mcp.
- If you used an API key: Rotate or revoke the key on the firecrawl.dev/app/api-keys page; you can also set a credit limit per key.
- For a local install: Remove the global npx package with npm and stop the Docker container; the documentation has no official removal steps.
Frequently asked questions
Do you need an API key for Firecrawl MCP?
No, but the scope narrows. The keyless endpoint offers only search, page reading and file parsing and has daily per-IP request and credit limits. The full tool list needs an account sign-in or an API key. The local npx package needs FIRECRAWL_API_KEY for the cloud API.
How much can I use on the free plan?
1,000 credits a month (about 1,000 pages or 500 searches), no card required. When credits run out, requests return HTTP 402; the free plan has no pay-as-you-go.
What is the daily limit of the keyless free tier?
The official documentation does not give the number: it says there are two daily limits per IP (request count and credits). It notes that keyless calls may be rejected from some networks.
How are credits spent, and what is expensive?
Page reading costs 1, search 2 per 10 results, crawling 1 per page. What makes it more expensive: JSON extraction +4/page, PDF +1/page, interact 2 to 7 per browser minute. If you set no limit on crawl, the default cap is 10,000 pages; setting a limit keeps credit burn in check.
Where does my data go?
Queries, URLs and extraction prompts go to Firecrawl's cloud API; the servers are in the US. Scraped pages are cached for 2 days by default. Zero data retention is only on Enterprise. If you run Firecrawl on your own server, the data stays on your infrastructure.
Can I host it on my own server?
The MCP server (MIT) runs on your computer with npx or Docker and can connect to your own Firecrawl server via FIRECRAWL_API_URL. The core Firecrawl engine is a separate repository licensed under AGPL-3.0.
Is it legal to scrape any site?
The official terms leave compliance with sites' terms of use and the law to you. robots.txt behavior is described only in the crawl documentation. Get legal advice before collecting personal data.
Related guides
Sources
- firecrawl-mcp-server README
- Firecrawl MCP documentation
- Firecrawl pricing
- Firecrawl billing documentation
- npm: firecrawl-mcp
Not clear in the official sources: Numbers for the keyless daily limit · map and interact credit rules (two official pages conflict) · robots.txt behavior for single-page reading · menu path and plan requirement for the ChatGPT and Claude native connector · admin permissions on team and enterprise accounts · whether paid plans cover commercial use · fix status of issues #306, #210, #311, #481 · availability restrictions in Turkey. Each is noted in the relevant section.