DeepScrape turns any website into a queryable MCP data source
Against Firecrawl's Fireround plugin, the author presents open-source DeepScrape, which uses an LLM once to generate self-healing CSS selectors and turns any site into an MCP source for Claude and Cursor.
Scraping companies are quietly turning into data companies. Firecrawl's new feature is the clearest tell: Fireround puts funding round data, $10B+ in weekly deals and 8,000+ investors, directly into ChatGPT and Claude through their plugin. It follows Firecrawl's $75 million Series B in August, raised specifically to build what they call Alexandria, a knowledge library for AI agents. The company that scrapes the web for you is starting to decide which slices of the web you get to query. That is a reasonable business, and also a ceiling. Fireround is fast because Firecrawl curated it ahead of time. Ask an agent to pull structured data from a site nobody at Firecrawl thought to index and you are back to raw extraction, paying an LLM to re-read the same page structure on every single call. DeepScrape takes the other path. Point it at a site once and it generates deterministic CSS selectors with an LLM, then reuses them on every later call, no model re-reading the page unless the layout actually changes, in which case it self-heals the selector instead of burning tokens from scratch. The LLM only does the expensive part once: finding, validating, and adapting the selector. Site-to-MCP takes that further and turns any website, not just a handful of curated verticals, into a structured source an agent can query directly inside Claude or Cursor. Curated libraries and generic self-healing extraction are both going to exist. I would rather my agent have both options than wait for someone else to decide which site was worth indexing. https://github.com/stretchcloud/deepscrape https://x.com/firecrawl/status/2110759634164220368
This post quotes another post: x.com/firecrawl/status/2110759634164220368