Fetch and Convert Web Content to Markdown

A Model Context Protocol server that fetches web content and converts it to clean, readable markdown using the Defuddle library, positioned as a drop-in replacement for the default MCP fetch server, which relies on Mozilla's Readability.js instead. The project's own motivation, drawn from a GitHub issue on the Defuddle repository, is that Readability.js performs noticeably worse than Defuddle at extracting content from certain modern sites - GitHub pages specifically called out as a weak spot - and that Defuddle's support for custom per-site extractors makes it a stronger foundation for a general-purpose web-fetching MCP server.

It exposes a single tool, fetch, which retrieves a URL and extracts its content as clean markdown (or raw HTML, if preferred). Parameters are url (required), max_length (optional, defaulting to 5000 characters, for capping how much content is returned), start_index (optional, defaulting to 0, for paginating through long pages by starting extraction at a given character offset), and raw (optional boolean, defaulting to false, to skip markdown conversion and return raw content instead). Beyond the cleaned content itself, the tool returns rich metadata: title, author, publication date, word count, domain, and processing time.

Three features summarize the value proposition: better content extraction (Defuddle strips webpage clutter while keeping the page title and key metadata intact), flexible output (markdown or raw HTML), and chunked reading (via start_index/max_length pagination for long pages that would otherwise be truncated or overwhelming in a single fetch).

Use this connector as a general-purpose replacement for the default fetch MCP server whenever an AI assistant needs to read and summarize web page content, particularly for sites where Readability.js-based extraction produces messy or incomplete results (GitHub pages are the documented example). It is a straightforward drop-in swap - same fetch-style interface, better extraction quality - rather than a fundamentally different tool, so existing fetch-based workflows can adopt it without redesigning how content requests are made.

A single fetch tool, functionally interchangeable with the default fetch MCP server's own tool signature, so a client already calling that server's fetch can point at this one without changing how it invokes it - only the extraction quality and the added metadata fields change.

{
  "mcpServers": {
    "defuddle-fetch": {
      "command": "npx",
      "args": ["-y", "defuddle-fetch-mcp-server"]
    }
  }
}

Install via npx defuddle-fetch-mcp-server and configure into Claude Desktop through an mcpServers entry in claude_desktop_config.json; can also be built and run locally from source with Node.js for development or contribution, and follows semantic versioning for releases published to the NPM registry via GitHub Actions.

Anyone using an MCP client like Claude Desktop who wants more reliable web-page-to-markdown extraction than the default Readability.js-based fetch server provides, especially when regularly fetching content from sites like GitHub where the default extractor performs poorly.