An agent that fetches web pages itself gets HTML: scripts, styles, menus and footers, often tens of thousands of tokens before the first useful sentence. Pages built by JavaScript come back nearly empty. Readmint gives an agent tools that return what a person would read.
The tools
Readmint is an MCP server at https://api-readmint.plxi.us/mcp, with three
tools:
| Tool | What the agent gets |
|---|---|
read_url |
A page's main content as Markdown, with its title, author, date and language. Optional: one part of the page by CSS selector, a character cap, links, images and metadata. |
preview_url |
A link's title, description, image, site name, icon and oEmbed details. |
map_site |
A site's page URLs, from its sitemaps and the links on a start page. |
Each tool call costs one page from your plan, like the REST call it mirrors. Failed reads are free.
Set it up
The docs, and the dashboard's API keys page once you create a
key, show the setup for Claude Code, Cursor and VS Code, plus a block of
instructions to paste into your agent's AGENTS.md or CLAUDE.md, such as:
To read a web page, call
read_urlinstead of fetching its HTML yourself.
Why not fetch pages directly?
- Context: Markdown of the main content is a fraction of the page's HTML.
max_charscaps it further, andtruncatedtells the agent there's more. - JavaScript: pages built by script are loaded in a real browser when they need it, without the agent having to know.
- Safety: the tools only fetch public addresses, so a prompt-injected link
to
localhostor a cloud metadata address gets nowhere. - Honest failures: bot checks, logins and HTTP errors come back as clear errors the agent can act on, not as a challenge page it might summarise.