Fetch MCP

What zcaceres/fetch-mcp MCP server does

The zcaceres/fetch-mcp MCP server exposes web-fetching operations through the Model Context Protocol. It can return a page as raw HTML, convert it to Markdown, or remove tags, scripts, and styles to produce plain text. A separate JSON operation returns the response from a URL as JSON.

For longer pages, the server supports character-based response controls. Callers can set max_length to limit returned content and start_index to begin at a later character position. Custom HTTP headers and a proxy URL can also be supplied per request.

Article-oriented extraction is available through fetch_readable. It uses Mozilla Readability to identify the main article and returns that content as Markdown while removing navigation, advertisements, and other page boilerplate. fetch_youtube_transcript retrieves captions or transcripts from YouTube and accepts a lang parameter, which defaults to en.

How it works

Each MCP tool requires a url. The common optional inputs are headers, max_length, start_index, and proxy. The default maximum response length is 5,000 characters unless changed through the request or the DEFAULT_LIMIT environment variable.

YouTube transcript retrieval uses yt-dlp when it is available. If that dependency is not available, the server attempts direct extraction from the YouTube page. The requested caption language is selected with the tool’s lang parameter.

The project includes protections against requests to private or localhost addresses and DNS rebinding. It also limits response bodies to reduce memory usage; the default maximum is 10 MB and can be changed with MAX_RESPONSE_BYTES.

Setup and configuration

Install the zcaceres/fetch-mcp MCP server by configuring an MCP client to run the published mcp-fetch-server package with npx. A direct configuration uses the mcp-fetch-server package as the command argument. DEFAULT_LIMIT sets the default character limit, with 0 disabling that limit, while MAX_RESPONSE_BYTES controls the maximum response body size in bytes.

The same project provides a CLI package command named mcp-fetch. Its commands include html, markdown, readable, txt, json, and youtube. CLI flags cover response length, starting character position, proxy selection, and YouTube caption language.

Tools and capabilities

• fetch_html: returns raw HTML.
• fetch_markdown: converts fetched content to Markdown.
• fetch_txt: returns text with HTML tags, scripts, and styles removed.
• fetch_json: returns a URL’s JSON response.
• fetch_readable: extracts primary article content as Markdown.
• fetch_youtube_transcript: retrieves YouTube captions or transcripts by language.
• Custom headers, proxy URLs, pagination-style offsets, and response limits.

Limitations and notes

The zcaceres/fetch-mcp MCP server is a fetch and extraction layer rather than a site-specific API client. The supplied material does not document authentication flows, JavaScript browser rendering, or support for named MCP clients. YouTube transcript results depend on captions or transcript data being available and on the server’s extraction path. Requests are subject to the server’s private-address protections and response-size limits.

The project is licensed under MIT. Its README also documents Bun-based development commands for installing dependencies, running watch mode, testing, and building, but those commands are for development rather than the normal MCP client setup.