Crawlbase says judge an MCP server by where the fetch comes from, and its own record cannot tell

Crawlbase says judge an MCP server by where the fetch comes from, and its own record cannot tell

Crawlbase has written a vendor comparison that is better than most because it names its own criterion up front. In The Best MCP Servers for Web Scraping: Nine Options Compared for 2026, Viktor Petrov lines up nine servers, Crawlbase, Bright Data, Firecrawl, Apify, Oxylabs, ScrapingBee, ScraperAPI, Jina, and Microsoft’s Playwright, and argues that tool count is the wrong axis. What matters is where the fetch originates: “The model only ever sees what the fetch returned.” A server that fetches from your own machine returns a block page where a plain script would, and “a fetch from your own IP will fail exactly where a plain script fails,” after which the model reasons confidently over the wrong content. Servers backed by a scraping network with proxies and rendering return the page.

The figures are the kind a comparison should carry, counts and prices rather than claims: Firecrawl registers 26 tools in its full profile, ScraperAPI 28, Crawlbase three, and the post turns that into a cost: “a server with dozens of tools spends more of the context window before any work starts than one with a handful.” Crawlbase’s own free tier is 5,000 requests with no card, a screenshot costs two credits, and “failed requests are not billed,” which it draws as a contrast with servers that charge for the block page. The advice is sound for any of the nine: “Test on 20 to 50 of your real URLs before you commit,” because vendor demos use cooperative sites, and treat an MCP config file holding an API key as a secret.

The catalog can place Crawlbase’s server but not grade it on the axis the post proposes. The Crawlbase provider page lists 5 API pages, and the MCP server’s three tools sit on two of them: crawl and crawl_markdown on the Crawlbase Crawling API, and crawl_screenshot on the Crawlbase Screenshots API. The MCP server dimension is lit, and the agentic access profile maps 11 operations, 4 of them acting. Where the fetch comes from is not something a contract declares, and it is not something the catalog measures, which is a fair point against the catalog and one the post makes without meaning to.

The Kin Score is 35.4, thin band. Discoverability carries it at 68.3 and contract quality at 48.4, with access clarity at 36.3, developer ergonomics at 32.1, operational transparency at 31.1, and contract governance at 0.0. The Agent Readiness score is 28.4, agent-aware. Error semantics is unlit, and that is the one dimension the post’s own argument turns on. A block page returned as a success is an error-semantics failure: the fetch succeeded and the content did not, and nothing in the response says so. Crawlbase is right that the model sees whatever the fetch returned. Its own contract does not yet describe how the fetch says when it returned the wrong thing.

← Cloudflare's MCP portals go GA, and its own record is already agent-native
Instrumentl's only public API is an MCP server, with eleven write tools and no contract →