Skip to content

Use it from Claude

The package includes an MCP server, and it gives an AI assistant four tools. The assistant fetches pages politely and searches the pages that it fetched before. It also reads a range of a page, and it checks a quote before it cites it.

Terminal window
claude mcp add polite-crawler -- npx -y polite-crawler mcp
Tool Input What it gives
crawl_fetch urls (1 to 20), render (auto, never or always), max_age (for example 24h) For each page: the outcome, the status, the document id, the title, the schema.org types and the capture ids.
crawl_find q, limit, kind Ranked passages from the stored pages, with their document ids.
crawl_read doc, from, length, user_provided A range of the text of a document, with its captures.
crawl_verify doc, quote The exact span of the quote and its position, or found: false.

crawl_find, crawl_read and crawl_verify send no request. They read the store on your disk.

A plain fetch tool goes back to the site for each question, and nothing stays. With polite-crawler, the assistant fetches a source one time, reads the stored copy many times, and cites the exact capture.

  • Rules for assistants. The MCP server obeys the robots.txt rules for Claude-User too, because an assistant asks for these pages. Test a robots.txt.
  • Content signals. When a page asks not to be AI input, the server keeps the page but holds back its text. When a person gave the URL directly, the assistant can pass user_provided, and the crawler records that choice.
  • Quotes from the source. crawl_verify finds the quote in the stored text and ignores only case, spaces and typographic punctuation. A citation thus rests on the text, not on the memory of the model.