The session resource is the part worth having: every query, page and screenshot with a timestamp, so a research pass can be retraced instead of reconstructed. If you skip the bundled research prompt, name the sources you want in the request — 'from Reuters or AP' rather than just 'news today' — because the quality of the search terms decides the quality of everything downstream.
A small web-research server built on Playwright: it runs the search, visits the pages, pulls out the content, and keeps the whole pass as a session you can refer back to.
- search_google runs a search and extracts the results
- visit_page loads a page, extracts its content, and can capture a screenshot in the same call
- take_screenshot captures the page currently open
- A research session kept as a resource: the queries run, the pages visited, the content extracted, the screenshots, and the timestamps for all of it
- Screenshots saved as MCP resources, attachable in the client afterwards
- An agentic-research prompt that steers a broad-then-narrow pass, prefers authoritative sources, and asks for every claim to carry its URL
Node.js 18 or later, plus Playwright, which comes in as a dependency. Nothing to authorize — no account and no API key. Point your client at npx -y @mzxrai/mcp-webresearch@latest over stdio. Verified on macOS, Linux and Windows. The author describes the code as pre-alpha, so treat it as such.
One command — npx -y @mzxrai/mcp-webresearch@latest