Scraping a URL you already have is the easy case; `search_and_scrape` means you can ask about something without knowing where the answer lives, and the model gets the pages rather than a list of links. Batch mode returning per-URL errors instead of failing wholesale is what makes a ten-page comparison survive one bad site.
The CrawlAPI SDK repository, which ships an MCP server alongside the JavaScript client. Point it at a URL and you get back markdown a model can read, with full JavaScript rendering — so pages that build themselves in the browser come back as content rather than an empty shell.
- `scrape_url` fetches one page and returns clean markdown
- `batch_scrape` handles up to 10 URLs in parallel; a failed URL returns an error field instead of failing the whole request
- `search_and_scrape` searches the web and scrapes the top results in one call
- Formats beyond markdown are available — html, text and structured — along with `waitFor` for JavaScript-heavy pages and a `timeout` defaulting to 30000 ms with a 60s maximum
- Scrape results carry metadata such as the page title
Node.js 18 or newer and a RapidAPI key — the free tier is 50 calls a day with no card. Clone the repository and run `npm install`, then point your client at `node` with the absolute path to `mcp-server.js` and `CRAWLAPI_KEY` in the environment. `CRAWLAPI_BASE_URL` overrides the API base if you need it. The same config works in Cursor, Windsurf and Continue; the server uses stdio transport.