DeepScrape: self-healing scraper with an MCP server
The author open-sources DeepScrape, which generates one CSS selector per page type, heals it on layout changes, and ships an MCP server so Claude and Cursor can call the scraper as a tool.
Firecrawl just gave AI agents free access to 270 million company and people records through something called Alexandria. 1,000 credits a month, build prospect lists, find work emails, research job history, all through a plugin that plugs straight into ChatGPT and Claude. It's a smart move and it tells you where the real cost center actually sits. The data itself was never the hard part. Pulling 270 million records together, keeping them current, and serving them without an agent burning its whole context window re-reading raw HTML on every single call, that's the part nobody wants to own, and it's exactly why people go looking for Firecrawl alternatives once their scraping bill shows up. That's the problem I built DeepScrape around. It generates a CSS selector once per page type, not once per request, so an agent isn't paying an LLM to re-parse the same layout a thousand times over. When the page changes, the selector heals itself instead of silently returning garbage. The LLM only gets called for the part that actually needs judgment: picking the right selector, validating it, adapting it when the site changes. Everything else stays deterministic. It also ships an MCP server, so Claude and Cursor can call a scraper as a tool directly, plus a site to MCP mode that turns any website into a structured source an agent can query on its own, with markdown, HTML, screenshots, tables, and change tracking coming out of the same crawl. The teams that figure out cheap, reliable web data first are going to ship agents the rest of us are still debugging. http://github.com/stretchcloud/deepscrape https://x.com/firecrawl/status/2112571818322116860
This post quotes another post: “Firecrawl free People Data: 270M+ company and people records for agents”