Give Claude, Cursor and Codex clean web pages with MCP

When an AI agent fetches a web page, most of what it gets back is not the page's point: navigation, cookie banners, related links, inline scripts. All of it fills the model's context window and some of it misleads the model. imread does the cleaning first and hands the agent only the article, as Markdown with headings, lists, tables, links and code kept.

In a hurry? Every setup below, with copy buttons, is on the MCP page.

One tool, no key

imread runs a Model Context Protocol server over Streamable HTTP at https://imread.org/api/mcp. It has one tool, extract_article, which takes a url and optionally a format (markdown or text), an offset and max_chars. There is no account and no API key.

Claude Code

claude mcp add --transport http imread https://imread.org/api/mcp

Codex CLI

codex mcp add imread --url https://imread.org/api/mcp

Cursor

Add this to ~/.cursor/mcp.json (or .cursor/mcp.json in a project):

{ "mcpServers": { "imread": { "url": "https://imread.org/api/mcp" } } }

VS Code

In .vscode/mcp.json:

{ "servers": { "imread": { "type": "http", "url": "https://imread.org/api/mcp" } } }

Claude on the web and desktop

Where your plan allows custom connectors, open Settings, then Connectors, choose Add custom connector and paste https://imread.org/api/mcp.

What the agent gets back

A long article arrives in parts, so it fits an agent's tool-output limit: each part ends with the offset to ask for next. imread also reads pages many fetchers can't: WeChat Official Account articles, Chinese and Japanese pages in legacy encodings such as GBK and Shift_JIS, and SEC EDGAR filings with their tables. Pages that build their text with JavaScript or sit behind a login come back with a readable error instead of junk.

Prefer plain HTTP? The same extraction is one GET request; the API docs have the details, and the MCP page keeps every setup above in one place.