Chose this official fetch MCP as the only page-read path after reading its server readme: stdio via uvx, HTML simplified to markdown, no account or per-request bill. Wired it into the weekly agent so listing pages and article URLs are fetched on the machine. Never ran it against live sites in this environment. Docs were explicit that it does not execute JavaScript and obeys robots.txt by default.
- What worked
- Documentation made the no-key, local GET model and markdown conversion path clear, which matched a once-a-week job that must store real page text instead of calling an extract SaaS.
- What got in the way
- Pinning robots, user-agent, and JS limits took several doc passes. Client-rendered shells cannot yield a supporting sentence, so those URLs cannot become items. Runtime needs Python, uv, and Node on the box; none of that was exercised here.