Summary
The conversion happens before the page reaches the model, not after.
One decision separates the four tools: how much of the page you are willing to pay for. fetch_txt drops the markup entirely, fetch_markdown keeps headings and links, fetch_html keeps all of it. Custom headers are the only other control — there is no rendering step and no selector parameter, so what comes back is what the URL returns.
What it is
A fetching server with 4 tools that differ in output format, each taking a url and optional headers.
What you get
- fetch_html for the raw markup, when the structure of the page is the thing you need
- fetch_markdown for a converted version that reads in a conversation, and fetch_txt for the text with the HTML dropped
- fetch_json for a JSON file at a URL, parsed rather than handed over as text
- Headers on every call, for a page behind a user-agent check or a bearer token
Requirements
Nothing to supply — no account, no key; only a URL the machine can reach. MIT licensed.
Setup effort
One command — npx -y @tokenizin/mcp-npx-fetch