Web-Data MCP Servers for AI Agents, Compared

#mcp #ai #webscraping #tooling

Originally published at AgentStack Notes. Last fact-checked: November 8, 2026.

Disclosure: This article was drafted with AI assistance and checked against each vendor's official documentation. We may receive a commission if you sign up through the ScraperAPI † or Firecrawl † links marked † below, at no extra cost to you. We have not tested these servers ourselves. Our applications to Apify, Bright Data and ScrapingBee are still pending, so links to those go to the official sites. See our Advertising Disclosure.

The Model Context Protocol (MCP) lets an AI client such as Claude Desktop, Cursor or VS Code call external tools. Most web-data vendors now ship an MCP server, so an agent can fetch a page or run a search without glue code. The servers differ in how you connect, how you sign in, and how much of the tool list is documented. This page compares five of them.

Note on experience: we have not run any of these servers in production. Facts below are from the vendor pages linked in the sources, read on November 8, 2026. Where a page did not say, we write "not stated" rather than guess.

1. At a glance

† marks a referral link.

Vendor	Remote endpoint (as documented)	Sign-in	Documented free allowance
Apify	`https://mcp.apify.com`	OAuth in the browser, or a bearer token; docs and Actor-search tools work without a token	Free plan: $5 of platform usage per month; the MCP page gives no separate MCP pricing
Firecrawl †	`https://mcp.firecrawl.dev/v2/mcp-oauth` (OAuth) or `https://mcp.firecrawl.dev/v2/mcp` (API key)	Keyless, browser sign-in, or API key	Keyless: Search, Scrape and Parse "within daily limits" (numbers not given). Free plan: 1,000 credits per month
Bright Data	`https://mcp.brightdata.com/mcp?token=YOUR_API_TOKEN` (from the project README)	Bright Data API token	5,000 requests per month, no card to start, no rollover
ScraperAPI †	`https://mcp.scraperapi.com/mcp`, or self-hosted with Python or Docker	Not stated on the page we read	Not stated for MCP; the regular plan is a 7-day, 5,000-credit trial, then 1,000 credits per month
ScrapingBee	`https://mcp.scrapingbee.com/mcp`	API key as the `api_key` query parameter	Same credits as the underlying API endpoint, no MCP surcharge; 1,000 free credits on the regular trial

2. Tools each server documents

Apify: enabled by default are search-actors, fetch-actor-details, call-actor, apify/rag-web-browser, apify/web-fetch, search-apify-docs, fetch-apify-docs and report-problem. Optional categories cover Actor runs, datasets, key-value stores, tasks and schedules. The rate limit is stated as up to 30 requests per second per user (HTTP 429 beyond that).
Firecrawl: firecrawl_scrape, firecrawl_map, firecrawl_search, firecrawl_parse, firecrawl_crawl, firecrawl_check_crawl_status, firecrawl_agent, firecrawl_agent_status, firecrawl_interact and firecrawl_interact_stop, plus groups for research, developer search, government search and monitors. The old Extract tool is described as deprecated.
Bright Data: 30 tools in total according to the project README, in groups named scrapers (6), browser (14), geo (2) and advanced_scraping (3). Every group also includes search_engine, scrape_as_markdown and ask_brightdata_assistant. You can limit the exposed tools with groups or tools parameters to keep the client's context small.
ScraperAPI: the MCP page says it supports ScraperAPI parameters such as geotargeting, premium proxies and JavaScript rendering. It does not list tool names.
ScrapingBee: HTML tools (get_page_text, get_page_html, extract_page_data, get_screenshot, get_file), search and AI tools (fast_search, get_google_search_results, ask_chatgpt, ask_gemini), Amazon and Walmart tools, YouTube tools, and get_scrapingbee_usage (limited to 6 calls per minute).

3. Which clients are named

Apify: Claude Desktop and VS Code extensions with MCP support; any client implementing the MCP specification should work. The page also points to a plugin for coding agents such as Claude Code, Cursor, GitHub Copilot and Codex.
Firecrawl: ChatGPT and Codex, Claude (including Claude Desktop and Claude Code), GitHub Copilot CLI and VS Code.
Bright Data (README): Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Gemini CLI, Zed and Warp, plus any client that supports local MCP servers.
ScrapingBee: Claude Desktop and Cursor via npx mcp-remote, ChatGPT through a custom app (plan limits apply), and any client that supports MCP over streamable HTTP.
ScraperAPI: not stated on the page we read.

4. Things to check before you connect

Where the key lives. ScrapingBee's page notes that the API key sits in the URL, so keep that config out of source control and rotate the key if it leaks. Bright Data's remote URL, as written in its README, also carries a token in the query string. Apify's default flow uses OAuth so the token stays out of the client config.
How many tools your client loads. A long tool list takes context. Bright Data documents group and tool filters for this reason.
Which meter runs. Apify bills platform usage, Firecrawl bills credits, Bright Data counts requests, ScrapingBee charges the same credits as the matching API call.
Concurrency and rate limits. ScrapingBee's hosted server has a shared global concurrency limit and may return HTTP 429; Apify states 30 requests per second per user.

5. A neutral way to choose

Try an MCP server without an account: Firecrawl documents a keyless mode with daily limits, and Apify's docs and Actor-search tools work without a token.
Most documented tool surface and filters: Bright Data lists 30 tools in named groups.
A specific site or marketplace: ScrapingBee lists Amazon, Walmart and YouTube tools; Apify's Actor catalog is its equivalent.
You already use an HTTP scraping API and want the same engine in your agent: ScraperAPI † and ScrapingBee both document MCP servers on top of their APIs.

6. What we could not confirm

ScraperAPI MCP: tool names, supported clients, authentication and credit usage.
Bright Data: the product page and the docs overview did not state the endpoint, tool count or authentication; those details come from the project README.
Firecrawl: credit cost per MCP tool (only Alexandria discovery is stated as free) and the numeric keyless daily limits.
Real-world reliability, latency and result quality of any of these servers; we have not measured them.

DEV Community