Guide

Firecrawl MCP: setup, tools, limits, and a hosted alternative

Firecrawl MCP setup, tools, auth, credit costs, and where the keyless tier stops, plus how datamint's hosted MCP server compares.

Firecrawl ships an official MCP server. Its README documents 27 tools, three hosted endpoints, and a keyless mode that works without an account. That breadth is useful. It also makes the setup decisions less obvious than they look, because the endpoint you choose, the auth mode, and the tool profile all change what your agent can do and what it costs.

Facts about Firecrawl here come from its MCP documentation and GitHub README, read on 20 October 2026. Credit rules and endpoints change, so check the source links before you commit a budget.

What the Firecrawl MCP server is

MCP is the protocol that lets an agent discover and call external tools. Firecrawl's server wraps its scraping and crawling API in that protocol, so a client such as Claude Code or Cursor can search, scrape, map, and crawl without custom glue code.

Firecrawl documents three hosted surfaces:

  • https://mcp.firecrawl.dev/v2/mcp is the keyless and API-key endpoint. Without a key it exposes firecrawl_scrape, firecrawl_search, and firecrawl_parse. With a bearer token it exposes the full tool set.
  • https://mcp.firecrawl.dev/v2/mcp-oauth is the OAuth endpoint. The user signs in through the browser, approves a team, and the client receives a token. No API key goes into the config.
  • https://mcp.firecrawl.dev/v2/mcp-search is a search-only endpoint with nine tools, aimed at clients that need search without the rest of the surface.

There is a local path too. The README documents npx -y firecrawl-mcp with FIRECRAWL_API_KEY in the environment, a global npm install, a local Streamable HTTP mode on port 3000, and a Docker image at ghcr.io/firecrawl/firecrawl-mcp-server:latest.

The OAuth and API-key endpoints reach the same full tool set. The difference is how the credential travels. OAuth suits a team because access can be reviewed and revoked from MCP settings. An API key suits a shared machine or a CI job because no browser flow is involved.

Connect it to Claude Code, Cursor, or Claude Desktop

The local npx configuration has the same shape across the clients Firecrawl documents. Put this in the client's MCP config:

{
  "mcpServers": {
    "firecrawl": {
      "command": "npx",
      "args": ["-y", "firecrawl-mcp"],
      "env": {
        "FIRECRAWL_API_KEY": "fc-YOUR_API_KEY"
      }
    }
  }
}

For a hosted connection, point the client at https://mcp.firecrawl.dev/v2/mcp and send Authorization: Bearer fc-YOUR_API_KEY. Clients that support remote MCP servers take a URL and a header rather than a command. Firecrawl documents Cursor, Windsurf, VS Code, and Claude Desktop, and its docs list native connectors for Claude and ChatGPT/Codex.

To evaluate without an account, use the same URL with no key. The server exposes scrape, search, and parse only. Crawl, map, agent, interact, and monitors need a credential.

Keep the API key out of the server URL and out of agent chats. MCP configs are often committed to repositories, and the README calls this out for a reason.

The tool surface

The full profile documents 27 tools. Reading the list is the fastest way to see whether Firecrawl fits your workload, because the surface is much broader than "scrape a page."

The core fetch tools are firecrawl_scrape, which returns Markdown, schema-based JSON, branding data, or other formats for one URL, and firecrawl_parse, which handles PDFs and other documents. Discovery is covered by firecrawl_map for URLs and firecrawl_search for ranked web results. firecrawl_crawl starts a multi-page crawl and firecrawl_check_crawl_status polls it.

Beyond that, the server exposes an autonomous firecrawl_agent and its status tool, firecrawl_interact for clicking, typing, and form filling in a page session, a monitor family for recurring page checks, research tools for papers, developer and government search, a tool catalogue search, and firecrawl_credit_usage.

The breadth is the product's main advantage. A coding agent can search documentation, crawl a site, parse a PDF, monitor a page for changes, and check its own credit balance without leaving the client. The tradeoff is that 27 tools is a large menu for a model to choose from. If your workload is one URL to clean JSON, most of that surface is irrelevant context.

The keyless profile is much smaller. Only firecrawl_scrape, firecrawl_search, and firecrawl_parse are exposed, and Firecrawl describes the daily limits as rate-limited without publishing a number.

What it costs

Firecrawl publishes some costs in the MCP README and leaves others to the plan.

A firecrawl_search call costs 2 credits. A first, usable firecrawl_search_feedback submission refunds 1 credit, with refunds capped at 100 credits per team per UTC day. Agent jobs take a maxCredits setting that defaults to 2,500, and a free request is capped at the same 2,500. Crawl results report creditsUsed.

The README does not publish a per-call credit cost for scrape, map, crawl, or parse. Those draw on the team's Firecrawl plan. A budget estimate therefore has to start from the pricing page or from observed creditsUsed values.

The keyless tier has no price because it has no account. The restriction is the tool list and the daily limit.

Where the keyless tier stops

Keyless access works for evaluation. Production needs an account.

It covers a single scrape, a search, and a document parse. A crawl, a map, an agent run, an interactive session, or a monitor all require an API key or an OAuth connection. The daily limit is not published, so a pipeline that depends on the free tier has no capacity guarantee.

If you are testing whether MCP is useful for your agent, three tools are enough to see the shape of the workflow. If you are testing whether Firecrawl can crawl 500 pages on a schedule, you need an account before the test starts.

Limits worth knowing before you build

The README is explicit about several constraints that affect architecture.

firecrawl_scrape handles one URL per call. firecrawl_parse handles one file per call. Bulk operations belong to the Firecrawl API, not the MCP tools. A crawl can produce more content than an agent's context window can hold, so the README recommends bounded depth and page limits. Agent jobs are asynchronous and can take minutes, with polling suggested every 10 to 30 seconds. firecrawl_interact completes one prompt or code turn per call, though a session persists through scrapeId. Under CLOUD_SERVICE=true, scrape is read-only and parse accepts uploads only.

These constraints fit a hosted tool surface designed around agent-sized calls. They matter when you map an MCP workflow onto a batch job, because the MCP layer is not a throughput interface.

How datamint's hosted MCP server compares

datamint takes the opposite trade. The server is hosted at https://mcp.datamint.xyz/sse and exposes five tools:

  • datamint_mint fetches one URL and returns clean Markdown or structured JSON.
  • datamint_extract returns JSON validated against a canonical schema or a custom JSON Schema.
  • datamint_crawl runs an asynchronous crawl with depth, page, and path filters.
  • datamint_crawl_status polls the crawl and returns pages.
  • datamint_estimate_cost calculates the credit and USD cost before a run.

Authentication is an API key, either as the api_key argument on a tool call or as DATAMINT_API_KEY in the environment. The hosted endpoint uses SSE, and the docs also cover a local stdio or container path.

Two differences matter in practice: the output and the credit model. Firecrawl returns Markdown, HTML, or JSON from its scrape formats. datamint's schema pass returns JSON that either validates against the schema you supplied or fails the request. That matters when the next step in the pipeline expects a fixed shape and you do not want to write a parser for each target.

On credits, datamint publishes 1 credit for a static fetch, 5 for a headless render, 15 for stealth routing, and 2 for a schema pass, with $1 equal to 1,000 credits. The free tier is 1,000 credits per month. datamint_estimate_cost prices a workload before you run it.

Firecrawl is stronger on search, map, autonomous agent runs, monitors, PDF parsing, research tools, and the keyless evaluation tier. If your agent needs to discover URLs, watch pages for changes, or work across papers and government sources, Firecrawl covers operations datamint does not.

datamint is stronger on typed output, a small tool surface, published per-page credit costs, PII stripping at ingestion, and configurable robots.txt compliance. If your workload is "take this URL, return validated JSON from a defended site," datamint is built for that narrower job. See the URL to Markdown API for the single-page path.

The two can sit in the same stack. Use Firecrawl for discovery and monitoring, then pass the URLs to datamint when you need validated extraction. That division of labor is normal and it does not require a migration. For the pricing-level comparison, see the Firecrawl alternatives page and the Firecrawl pricing guide.

Connecting datamint's server

Cursor and other clients that accept a remote URL:

{
  "mcpServers": {
    "datamint": {
      "url": "https://mcp.datamint.xyz/sse"
    }
  }
}

Claude Code:

claude mcp add --transport sse datamint https://mcp.datamint.xyz/sse

Claude Desktop uses mcp-remote to bridge the hosted SSE endpoint to stdio:

{
  "mcpServers": {
    "datamint": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.datamint.xyz/sse"]
    }
  }
}

The first tool call needs an API key. You can create one on the dashboard, and the free tier includes 1,000 credits. Full setup details and the tool schemas are in the datamint MCP documentation.

Which one should you use?

Pick Firecrawl when the job needs discovery, search, crawling, monitoring, or an autonomous agent, and when a keyless evaluation matters. Pick datamint when the job is extract-this-page-into-this-schema and the target may be defended. Use both when the workflow starts with discovery and ends with structured data.

An MCP server is not a drop-in replacement for another one. The tool names, auth model, output shape, and credit units are different. Treat a migration as a small integration project. A config swap will not cover it.

FAQ

Is Firecrawl MCP free?

There is a keyless mode with firecrawl_scrape, firecrawl_search, and firecrawl_parse, subject to unpublished daily limits. The full tool surface requires an API key or an OAuth connection tied to a Firecrawl plan.

Does Firecrawl MCP work with Claude Code and Cursor?

Yes. Firecrawl documents Cursor, Windsurf, VS Code, and Claude Desktop, and its docs list native connectors for Claude and ChatGPT/Codex. Claude Code can use the local npx config or a remote MCP connection with a bearer token.

What is the difference between the Firecrawl MCP endpoints?

/v2/mcp is the keyless and API-key endpoint. /v2/mcp-oauth is the OAuth account endpoint. /v2/mcp-search is a search-only endpoint with nine tools.

How many tools does Firecrawl MCP expose?

The full profile documents 27 tools. The keyless profile exposes three, and the search-only endpoint exposes nine.

Is there a Firecrawl MCP alternative?

datamint runs a hosted MCP server with five tools and schema-validated output. Community Firecrawl-compatible servers exist, but most wrap the same Firecrawl API and therefore use the same plan and credits. Compare the tool list and the output shape before switching.

Can datamint's MCP server crawl a site?

Yes. datamint_crawl takes a seed URL, a depth limit, a page budget, path filters, and an optional webhook. datamint_crawl_status returns the pages and progress. It is asynchronous, like Firecrawl's crawl tool.

Sources

All three were read on 20 October 2026. Firecrawl's endpoints, tool list, and credit rules can change, so verify them against the current docs before relying on a number in production.

This guide covers retrieving publicly accessible data. Check the target's robots.txt and terms before crawling it, and do not use either server to access authenticated or paywalled content.

datamint returns schema-validated JSON and LLM-ready Markdown from bot-defended sites through one API call, with routing across transport tiers handled on our side.

All guides