Fetch and convert web pages to markdown for LLM consumption

Fetch MCP Server

A Model Context Protocol server that provides web content fetching capabilities. This server enables LLMs to retrieve and process content from web pages, converting HTML to markdown for easier consumption.

Source: https://github.com/modelcontextprotocol/servers/tree/main/src/fetch

Requires MCP Python SDK 2.x (mcp>=2.3,<3).

The fetch tool will truncate the response, but by using the start_index argument, you can specify where to start the content extraction. This lets models read a webpage in chunks, until they find the information they need.

The markdown of an HTML page opens with the page's title as a # heading (its og:title, else its <h1> when the <title> holds it, else its <title> or <h1>), because the extraction drops a heading that repeats the title. It is not added when the markdown already opens with it, to a raw result, or to a page that failed to simplify.

Available Tools

• fetch - Fetches a URL from the internet and extracts its contents as markdown.
  • url (string, required): URL to fetch
  • max_length (integer, optional): Maximum number of characters to return (default: 5000)
  • start_index (integer, optional): Start content from this character index (default: 0)
  • raw (boolean, optional): Get raw content without markdown conversion (default: false)
  • timeout_ms (integer, optional): HTTP request timeout for this call, in milliseconds, from 100 to 600000 (default: the server's timeout, 30000 unless configured; see Customization - Timeout)
  • Annotations: readOnlyHint: true, destructiveHint: false, idempotentHint: true, openWorldHint: true (it only reads, and it reaches arbitrary hosts on the internet)

Prompts

• fetch
  • Fetch a URL and extract its contents as markdown
  • Arguments:
    • url (string, required): URL to fetch

Errors

• A failed fetch (a robots.txt refusal, an unreachable page, an HTTP error status, a refused private address, an unusable proxy setting) and arguments that do not match the tool's schema are returned as a tool result with isError: true, whose text starts Error executing tool fetch: and says what went wrong, so the model can read it and decide what to do.
• Calling a tool the server does not have is a JSON-RPC protocol error, -32602 (Unknown tool: <name>), as the specification requires. So is getting a prompt it does not have (Unknown prompt: <name>), or the fetch prompt without its url (URL is required) or with a url that is not a valid URL (Invalid URL '<url>': <reason>).
• When the fetch prompt's fetch fails, the prompt still returns: its message holds the reason the fetch failed.

The server is built on the MCP Python SDK's high-level MCPServer: the tool and the prompt are decorated functions, and their type hints and field descriptions are the input schema and the prompt's arguments.

Content negotiation

The tool and the prompt ask for markdown first, with the request header Accept: text/markdown, text/html;q=0.9, */*;q=0.8. A site that serves markdown through content negotiation sends it directly, and a text/markdown response is returned as-is: it is not simplified and carries no raw-content notice. A site that ignores the header sends what it always did, and HTML is simplified to markdown as before. A raw fetch keeps the default Accept: */*, so it gets whatever the site serves by default (for a web page, its HTML), unsimplified; a markdown response it receives anyway is returned as-is too.

Protocol versions

The server speaks both eras of the Model Context Protocol over stdio, with nothing to configure: a 2026-07-28 client (discovered with server/discover, no session) and a 2025-11-25 client (the initialize handshake). The client's first request decides which. Tools and prompts behave the same in both. Methods the 2026-07-28 specification removes, such as ping, are answered only on a 2025-11-25 connection.

Installation

Optionally: Install node.js, this will cause the fetch server to use a different HTML simplifier that is more robust. Its JavaScript dependencies are then installed with npm on first use, which contacts the npm registry. To keep the server from ever starting Node.js, see Customization - HTML extraction.

Using uv (recommended)

When using uv no specific installation is needed. We will use uvx to directly run mcp-server-fetch.

Using PIP

Alternatively you can install mcp-server-fetch via pip:

pip install mcp-server-fetch

After installation, you can run it as a script using:

python -m mcp_server_fetch

Configuration

Configure for Claude.app

Add to your Claude settings:

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

{
  "mcpServers": {
    "fetch": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "mcp/fetch"]
    }
  }
}

{
  "mcpServers": {
    "fetch": {
      "command": "python",
      "args": ["-m", "mcp_server_fetch"]
    }
  }
}

Configure for VS Code

For quick installation, use one of the one-click install buttons below...

For manual installation, add the following JSON block to your User Settings (JSON) file in VS Code. You can do this by pressing Ctrl + Shift + P and typing Preferences: Open User Settings (JSON).

Optionally, you can add it to a file called .vscode/mcp.json in your workspace. This will allow you to share the configuration with others.

Note that the mcp key is needed when using the mcp.json file.

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

{
  "mcp": {
    "servers": {
      "fetch": {
        "command": "docker",
        "args": ["run", "-i", "--rm", "mcp/fetch"]
      }
    }
  }
}

Customization - robots.txt

By default, the server will obey a websites robots.txt file if the request came from the model (via a tool), but not if the request was user initiated (via a prompt). Every redirect hop of a tool fetch is checked against the robots.txt of the host it leads to, the same way as the requested URL. If the robots.txt request fails to connect, or is answered with 401, 403 or a server error (5xx), the tool refuses the fetch rather than reading the response as rules (RFC 9309 treats a 5xx robots.txt as complete disallow); any other 4xx is treated as no rules. This can be disabled by adding the argument --ignore-robots-txt to the args list in the configuration.

Customization - Private addresses

By default, the server refuses any request, including a redirect hop or a robots.txt fetch, whose host resolves to an address that is not globally routable: loopback (127.0.0.0/8, ::1), private (10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, fc00::/7), link-local (169.254.0.0/16, fe80::/10), shared address space (100.64.0.0/10), multicast, unspecified, documentation and the other IANA special-purpose ranges. An IPv4-mapped (::ffff:0:0/96) or well-known NAT64 (64:ff9b::/96) address is judged by the IPv4 address it carries; a deprecated IPv4-compatible address (::/96, such as ::127.0.0.1) is refused whatever IPv4 it carries. A hostname is refused if any one of its addresses is in those ranges. The client receives an error naming the refused address.

To fetch from local or internal hosts (a development server, an intranet page), add the argument --allow-private-ips to the args list in the configuration.

Customization - Allowed hosts

By default the server may fetch from any host (subject to the private-address guard above). To restrict it to the hosts you name, add --allowed-hosts followed by one or more entries to the args list (the argument can also be repeated). Each entry is one of:

• an exact host name, such as docs.python.org, which matches that host only, not its subdomains;
• *. followed by a domain, such as *.github.com, which matches every subdomain of it at any depth (api.github.com, a.b.github.com) but not github.com itself: list both to allow both;
• an IP address, such as 203.0.113.7 or 2001:db8::7, which matches a URL that uses that address literally.

Matching ignores case and a trailing dot, and an internationalized name may be given in either its Unicode or its xn-- form. An entry names a host only, so it carries no scheme, port or path. The list is checked before every request the server sends, by the host name in the URL rather than by the addresses it resolves to: the requested page, its robots.txt check, and every redirect hop of either, from the tool and from the prompt alike. A request to any other host is refused before it is sent, and the client receives an error naming the refused host.

The allowlist and the private-address guard are separate checks, and a request must pass both. Listing an internal host, such as localhost or 10.0.0.5, does not by itself let the server reach it: that also needs --allow-private-ips, which in turn does not lift the allowlist.

{
  "mcpServers": {
    "fetch": {
      "command": "uvx",
      "args": ["mcp-server-fetch", "--allowed-hosts", "docs.python.org", "*.wikipedia.org"]
    }
  }
}

Customization - User-agent

By default, depending on if the request came from the model (via a tool), or was user initiated (via a prompt), the server will use either the user-agent

ModelContextProtocol/1.0 (Autonomous; +https://github.com/modelcontextprotocol/servers)

or

ModelContextProtocol/1.0 (User-Specified; +https://github.com/modelcontextprotocol/servers)

This can be customized by adding the argument --user-agent=YourUserAgent to the args list in the configuration.

Customization - Timeout

The HTTP request for the page waits up to 30 seconds (30000 ms) by default: to connect, and then for each read of the response. It bounds each of those separately, not the whole download, so a large page that keeps arriving is not cut off. A request that times out fails with a tool error saying how long it waited.

The timeout can be set at three levels, the most specific winning:

1. The fetch tool's timeout_ms argument, for one call (for example 120000 for a large download, or 5000 to fail fast).
2. The argument --timeout-ms=60000 in the args list in the configuration, for every call.
3. The FETCH_TIMEOUT_MS environment variable (for example "env": {"FETCH_TIMEOUT_MS": "60000"} in the configuration), for every call when --timeout-ms is not given.

The server-wide value also applies to the fetch prompt, and is the default the tool advertises for timeout_ms. Every level takes a whole number of milliseconds from 100 to 600000 (10 minutes); the server refuses to start with a value outside that range, and the tool refuses such a timeout_ms. This is the HTTP request timeout only: the robots.txt check before an autonomous fetch keeps its own 5-second timeout, and HTML extraction is not covered.

Customization - Proxy

The server can be configured to use a proxy by using the --proxy-url argument.

Without --proxy-url, the server uses the proxy in the standard HTTP_PROXY, HTTPS_PROXY and ALL_PROXY environment variables (honoring NO_PROXY). Both http(s):// and SOCKS5 proxies are supported (SOCKS support is installed with the server). A socks:// proxy URL, as many desktop proxy settings export it, is treated as socks5://. A proxy setting the server cannot use makes each fetch fail with a tool error that says so.

Customization - HTML extraction

HTML is converted to Markdown with readabilipy. When Node.js 14 or newer is on the PATH, the server runs Mozilla's Readability.js (readabilipy's ExtractArticle.js) in a Node.js subprocess; otherwise it uses readabilipy's pure-Python extractor, which keeps headings and text but drops link targets. The subprocess runs off the server's event loop, so other requests are answered while it works, and under a 30 second timeout: a Node.js that hangs is killed, together with anything it started, and that page is extracted with the pure-Python extractor instead. If Readability.js fails in any other way, the same fallback is used. Every reason for a fallback is written to stderr (the server's log), never to stdout, which carries the protocol.

Readability.js's dependencies are installed with npm on first use, which reaches the npm registry. readabilipy ships only the package.json for its JavaScript (@mozilla/readability, jsdom, minimist). The first time the server extracts HTML with Node.js on the PATH and those packages missing, it runs npm install in readabilipy's javascript directory, inside that tool call, so:

• registry.npmjs.org belongs in the egress allowlist of a deployment that restricts outbound traffic, unless the packages are installed ahead of time (below) or Readability.js is turned off. Without it the install fails and the server uses the pure-Python extractor.
• The JavaScript that runs is resolved from readabilipy's unpinned version ranges when it is installed, so pinning mcp-server-fetch does not pin it.
• The install runs at most once per server process, under a 120 second timeout. Its output goes to stderr. If it fails, or Node.js is missing or older than 14, the server uses the pure-Python extractor until it is restarted.

To install the packages ahead of time (in a container image, say), run npm install in readabilipy's javascript directory from the Python environment the server runs in; the server then finds them and installs nothing:

cd "$(python -c "import os, readabilipy; print(os.path.join(os.path.dirname(readabilipy.__file__), 'javascript'))")"
npm install --no-audit --no-fund

Readability.js discards hidden elements. Pages rendered progressively on the server often send a visible loading shell, then the real content in a container hidden (display: none, visibility: hidden, hidden) until scripts run in the browser. When a hidden element other than navigation or similar page furniture holds more text than Readability.js extracted, the server unhides it and runs Readability.js again, keeping that result only if it has more text. Other pages are extracted once, exactly as before.

To always use the pure-Python extractor, add the argument --no-readability-js to the args list in the configuration. The server then never starts Node.js (or npm), whether or not it is installed, and never contacts the npm registry.

Windows Configuration

If you're experiencing timeout issues on Windows, you may need to set the PYTHONIOENCODING environment variable to ensure proper character encoding:

{
  "mcpServers": {
    "fetch": {
      "command": "uvx",
      "args": ["mcp-server-fetch"],
      "env": {
        "PYTHONIOENCODING": "utf-8"
      }
    }
  }
}

{
  "mcpServers": {
    "fetch": {
      "command": "python",
      "args": ["-m", "mcp_server_fetch"],
      "env": {
        "PYTHONIOENCODING": "utf-8"
      }
    }
  }
}

This addresses character encoding issues that can cause the server to timeout on Windows systems.

Readability.js works on Windows as it does elsewhere: the server finds Node.js and npm (npm.cmd) on the PATH, and installs Readability.js's dependencies on first use as described in Customization - HTML extraction.

Debugging

You can use the MCP inspector to debug the server. For uvx installations:

npx @modelcontextprotocol/inspector uvx mcp-server-fetch

Or if you've installed the package in a specific directory or are developing on it:

cd path/to/servers/src/fetch
npx @modelcontextprotocol/inspector uv run mcp-server-fetch

Development

From src/fetch:

uv sync --frozen --all-extras --dev
uv run --frozen pytest

The tests drive serve() through an MCP ClientSession in-process (and, in tests/test_eras.py, through a Client as a 2026-07-28 and as a 2025-11-25 client), with HTTP mocked, so they need no network. The HTML extraction tests need Node.js and npm on PATH (the server uses Readability.js when Node is present and a pure-Python fallback when it is not), and the first run installs Readability.js's dependencies from the npm registry.

Coverage, measured per file with branch coverage:

uv run --frozen pytest --cov --cov-report=term-missing --cov-report=json

This writes coverage.json (git-ignored). The command reports coverage but does not enforce a threshold; the target is at least 90% of lines and 90% of branches for every file under src/mcp_server_fetch.