multifetch-mcp

A zero-config MCP server that lets Claude fetch any URL and search the web, with content returned as clean Markdown.

• Primary backend: Jina Reader (r.jina.ai + s.jina.ai)
• Automatic fallback: markdown.new (URL → Markdown) and markdown.new/search (web search via Serper.dev)
• Second fallback for fetch: defuddle.md (URL → Markdown with YAML frontmatter)
• No API key required for fetch (r.jina.ai is free). search uses the free markdown.new provider by default; set JINA_API_KEY to route search through Jina too and to raise fetch rate limits.
• Distributed via npm — runs with a single npx command.

Install

Claude Desktop

Open claude_desktop_config.json and add:

{
  "mcpServers": {
    "multifetch": {
      "command": "npx",
      "args": ["-y", "multifetch-mcp"]
    }
  }
}

Then restart Claude Desktop. That's it.

Claude Code

Bash / zsh / Git Bash:

claude mcp add multifetch -- npx -y multifetch-mcp

PowerShell (needs the --% stop-parsing token, otherwise PowerShell consumes the -- separator before it reaches claude):

claude mcp add multifetch --% -- npx -y multifetch-mcp

Windows CMD:

claude mcp add multifetch -- npx -y multifetch-mcp

If both forms fail in your shell, edit ~/.claude.json directly (see Manual config below).

Manual config

If the claude mcp add command misbehaves in your shell, you can edit ~/.claude.json directly. Find the mcpServers object (create it if missing) and add:

"mcpServers": {
  "multifetch": {
    "type": "stdio",
    "command": "npx",
    "args": ["-y", "multifetch-mcp"]
  }
}

Optional: higher rate limits

Set JINA_API_KEY (get a free one at jina.ai) to bypass the anonymous rate limit:

{
  "mcpServers": {
    "multifetch": {
      "command": "npx",
      "args": ["-y", "multifetch-mcp"],
      "env": {
        "JINA_API_KEY": "jina_..."
      }
    }
  }
}

Tools

fetch

Convert any URL to clean Markdown.

| Param | Type | Description | |---|---|---| | url | string (required) | Absolute URL to fetch | | format | markdown | frontmatter | html | text | Output format (Jina only). Default markdown | | no_cache | boolean | Bypass cached version | | target_selector | string | CSS selector to extract content from (Jina only) | | timeout | integer (1–180) | Max seconds to wait (Jina only) | | max_tokens | integer (≥500) | Truncate response to N tokens (Jina only) | | engine | auto | browser | Force a fetching engine. Use browser for JS-heavy sites |

Behavior: tries r.jina.ai first. If it fails (network / 4xx / 5xx / rate limit), falls back to markdown.new/<url>, then to defuddle.md/<url>. Returns a clear error only if all three providers fail.

search

Search the web and get the top results as Markdown.

| Param | Type | Description | |---|---|---| | query | string (required) | Search query | | site | string[] | Restrict to specific domains (Jina only) | | n | integer (1–5) | Number of results (markdown.new fallback). Default 3 | | no_cache | boolean | Bypass cached SERP (Jina only) |

Behavior: Jina's search endpoint (s.jina.ai) requires JINA_API_KEY — without one it returns 401. So by default (no key) search goes straight to markdown.new/search/<query> (powered by Serper.dev, top N pages) without a wasted Jina round-trip. When a key is set, it tries s.jina.ai first and falls back to markdown.new on any error.

Example prompts

Fetch https://example.com and summarize it.

Search the web for "MCP protocol" and give me the highlights from the top results.

Fetch https://news.ycombinator.com using the browser engine (JS-heavy site).

Develop locally

npm install
npm run build
node dist/index.js   # starts MCP server on stdio

To wire your local build into Claude Desktop, point the command at the absolute path of dist/index.js:

{
  "mcpServers": {
    "multifetch": {
      "command": "node",
      "args": ["C:\\path\\to\\multifetch\\dist\\index.js"]
    }
  }
}

Smoke-test with the official MCP inspector:

npx @modelcontextprotocol/inspector node dist/index.js

License

MIT