DonSeTch MCP Server: Web Fetch, Search and Crawl for AI Agents

• Reads a web page or PDF and returns clean Markdown

  Reads one URL, or up to 12 at once, and returns the page as Markdown. It can pull out only the part you ask about, check whether a page mentions something, or show what changed since last time. If a site blocks plain requests it switches to a headless browser, and it can click and type on a page when you pass steps. Scanned PDFs are read with OCR.

  Try: “Fetch https://docs.python.org/3/library/asyncio.html and show me only the section on task groups.”

  Network · web_fetch

• Searches the web with no API key and returns ranked links

  Searches many engines at once and ranks the results on your machine. You get links, titles and short snippets, not full pages. It can aim at GitHub, Wikipedia, Hacker News, Scholar or news, and it uses a search provider you added a key for if you set one up.

  Try: “Search for Rust async runtime benchmarks from this year and give me the top five links.”

  Network · web_search

• Walks a site from a starting page and collects its pages

  Starts at one page and follows links on the same site, optionally ranking pages by a topic you give. It can return just a list of URLs or the page contents. Big crawls stop at limits you set and give you a token to resume. It respects robots.txt by default.

  Try: “Crawl docs.python.org for pages about asyncio, up to 20 pages.”

  Network · web_crawl

• Takes a picture of a page in a real browser

  Opens the page in a headless browser and gives back a PNG image. The program also saves the same image in a screenshots folder in its cache on your machine. Local and private addresses are blocked like in fetch.

  Try: “Take a full-page screenshot of https://example.com/pricing.”

  Write · web_screenshot

Watch out for

• It can click and type on web pages

  web_fetch can run page steps such as click, type and press before reading. Avoid pointing it at logged-in or account pages you would not act on yourself. github.com

• Install downloads a program from GitHub

  The npm package fetches a prebuilt binary from GitHub Releases when you install it. The README says the download is checked with SHA256. github.com

• web_screenshot saves images on your computer

  Each capture is also written to a screenshots folder in the program's cache. Full-page captures can be several MB. github.com

• It works around bot walls and skips robots.txt on fetch

  It tries to look like Chrome and escalates to a browser when blocked. Fetch does not check robots.txt (crawl does), so follow each site's terms yourself. github.com

• It remembers pages it has read

  The README says fetches are fingerprinted and cookies persist, so later reads can report changes. Clear its cache if you fetched pages you want forgotten. github.com

• Optional unlocker sends URLs to Bright Data

  If you add an unlocker key, blocked fetches go through Bright Data's paid Web Unlocker. Without a key nothing is sent there. github.com

Security advisories