Firecrawl MCP: first connect the no-key hosted entry for search and scrape

Who it is for

People who need an MCP-compatible agent to search, scrape, and interact with the live web, and to get clean, model-ready context. It fits operators who can first connect the official hosted entry, accept that the no-key free tier only opens scrape / search / parse with rate limits, and know that crawl, map, and agent still need a key. It does not fit people who want to put a key into the server URL or the chat window, nor people who treat "connected" as unlimited whole-site crawling. Official repository: https://github.com/firecrawl/firecrawl-mcp-server

First try in 3 minutes

Take only one official path: first connect the hosted no-key free tier. Do not install local npx first, and do not configure a key first.

Prerequisite: the client can add a remote MCP. On the no-key tier scrape, search, and parse work but are rate-limited; crawl, map, and agent still need a key.

Install and config location: in the client, write the server address as the hosted entry, and do not append any key query parameter:

https://mcp.firecrawl.dev/v2/mcp

Low-risk action: after a successful connection you should see only 3 tools: firecrawl_scrape, firecrawl_search, firecrawl_parse. First call firecrawl_search, using a topic you already have; do not invent a website address that never appears in the evidence. If you must verify scrape, you may only scrape an address already in the evidence such as the official repository page https://github.com/firecrawl/firecrawl-mcp-server . Do not call crawl / map / agent, which need a key, on the first try.

Success signal: the client lists those 3 tools, search or scrape returns structured results, or a clear rate-limit note, not a demand to stuff a key into the URL. If authentication fails, first check whether you spliced a key into the address.

Use cases

Case 1: no-key retrieval or single-page scrape. This matches the source line search, scrape, and interact with the live web and the hosted entry exposing only firecrawl_search / firecrawl_scrape / firecrawl_parse. Use search for a batch of page summaries, or scrape to collect a single URL into clean structured data. The evidence says the no-key tier is rate-limited and does not write search query syntax or all scrape parameters, so you can only confirm the layer "clean context returned, or a rate-limit error".

Case 2: deeper capability after a key. A full default config registers 25 tools (including feedback tools); the source names crawl, map, agent plus "Deep research with autonomous agent" and "Interact with pages — click, navigate, and operate". These visit the live web; click and navigate have side effects. To take this path, you must configure Authorization: Bearer <FIRECRAWL_API_KEY> in the client's security header or key settings; never put the key in the server URL or the conversation. The evidence gives no parameter table for these 25 tools and no placeholder-free local npx start command, so do not invent a local install string.

Boundaries

The no-key hosted entry opens only 3 tools, and scrape / search / parse are rate-limited. The full tier's 25 tools, plus crawl / map / agent, sit outside this free boundary.

The source also has a search-only entry https://mcp.firecrawl.dev/v2/mcp-search The excerpt is truncated at "exposes a fixed 6", so you cannot assert the names of those 6 tools.

Never put an API key in the server URL or the agent conversation; put it only in the client's key management. readOnlyHint also does not mean zero risk: search and scrape both visit the live web and consume hosted quota.

The evidence gives no fixed client list. The local npx example in the excerpt carries a placeholder key and cannot be treated as a copyable official install command. For self-hosting or SSE, the source only lists capability names, with no step-by-step config; do not invent the missing parts.