MCP fetch server 403: the causes, and which server reads the page

The first cause is the request. The server sends the user agent “ModelContextProtocol/1.0 (Autonomous; +https://github.com/modelcontextprotocol/servers)”. A site that blocks bots needs no more than that line. A site that looks deeper sees the TLS handshake and the headers of a Python HTTP library, sent over HTTP/1.1. No browser sends those.

The second cause is a rule of the server. For a request that comes from the model, it reads robots.txt first. When robots.txt says no, the model gets an error and never sees the page.

{
  "mcpServers": {
    "fetch": { "command": "uvx", "args": ["mcp-server-fetch", "--ignore-robots-txt"] }
  }
}

The flags help on a few sites: 41 or 42 pages read, against 39. They do not change the handshake. For the other sites you need a server that sends a browser request, or a server that drives a real browser.

Anyfetch is a stdio MCP server, like the reference server. Change one entry in the MCP file of your client. With Node.js on the machine, there is nothing to install.

{
  "mcpServers": {
    "fetch": { "command": "uvx", "args": ["mcp-server-fetch"] }
  }
}

{
  "mcpServers": {
    "anyfetch": { "command": "npx", "args": ["-y", "@aliasmode/anyfetch", "mcp"] }
  }
}

Do you use Claude Code, Codex, Gemini CLI, or Cursor? The Anyfetch installer adds the server to each of them, so you do not edit a file.

No server in this test read every site. Anyfetch missed 15. A page behind a CAPTCHA, a login, or a subscription stays closed for a fetch server, and Anyfetch reports it with an error code such as challenge_required.

A task that needs a login or a click needs a real browser with a saved session. AliasMode is our free, open-source browser for that work, and its MCP server gives an agent profiles that stay logged in.