Because every query, page and screenshot lands in one accumulating session, a long piece of research keeps its trail — you can ask what was already read and get a list rather than a guess, and the bundled agentic-research prompt pushes for a URL beside every claim. Set expectations accordingly: the author calls the code pre-alpha, and steering it toward named sources ("from Reuters or AP") gets noticeably better material than a bare topic.
A three-tool browser research server. It drives Playwright to run a Google search, open a result and pull its content out, and it exposes the screenshots and the running session as MCP resources.
- search_google runs a query and extracts the results, so the model works from a live result page rather than what it remembers
- visit_page opens a URL and extracts the page content, with takeScreenshot capturing the page at the same time
- take_screenshot captures the current page on its own
- Screenshots are saved as MCP resources and reachable in Claude Desktop through the paperclip icon
- A research session accumulates the queries, the pages visited, the content extracted, the screenshots and their timestamps
- An agentic-research prompt ships with the package: broad searches first, authoritative sources preferred, direction refined as findings come in, and every claim cited with its URL
No account and no key. Node.js 18 or higher, and it starts as npx -y @mzxrai/mcp-webresearch@latest; Playwright comes with it as a dependency, so the browser it drives is installed for you. macOS is the platform the author lists as verified — Linux is not yet ticked. MIT.
One command — npx -y @mzxrai/mcp-webresearch@latest