Delv
Official (Anthropic)Active· 12d★ 4.3by Anthropic

Fetch

A simple HTTP fetch tool for Claude - read any public URL and get the cleaned text back. The no-frills alternative to Puppeteer for static pages.

View repo
A+
Safety & Trust

Delv Safety Grade: A+

Score 93/100 · assessed 2026-04-22

Maintainer95
Permissions95
Supply chain90
Transparency92
Incidents100

Fetch is an official Anthropic MCP server that performs HTTP GET requests and returns cleaned text content from public URLs. It's maintained in the official modelcontextprotocol/servers repository with active oversight from Anthropic. The tool has a narrow, well-scoped permission surface: it makes outbound HTTP requests to user-specified URLs and returns text, nothing more. No filesystem access, no shell execution, no credential handling. Distribution via uvx (Python package installer) is standard and verifiable. The codebase is fully open source with clear documentation. The main risk is SSRF-style attacks where Claude might be tricked into fetching internal network resources, though this is a protocol-level concern rather than an implementation flaw. For reading public web content, this is as safe as MCP servers get.

Lethal Trifecta (prompt-injection exposure)

ONE OF THREE
Private dataNo
Reads secrets, credentials, private files
Untrusted inputYes
Ingests web pages, PRs, issues, emails
External commsNo
Can send data outbound

Fetches arbitrary URLs — attacker can serve prompt-injection payloads. No private data or persistent outbound path by itself.

Green flags

  • Official Anthropic server with active maintenance
  • Read-only HTTP GET operations, no write capabilities
  • No filesystem, shell, or credential access required
  • Open source with clear documentation in main MCP repo
  • Simple install via standard Python package tooling (uvx)

Red flags

  • Could be tricked into fetching internal network URLs (SSRF vector)
  • No apparent rate limiting on outbound requests
  • Fetches arbitrary user-specified URLs without domain allowlist

Permissions requested

Outbound network
Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

Install

uvx mcp-server-fetch
Paste into your host's MCP config:
{
  "fetch": {
    "command": "uvx",
    "args": ["mcp-server-fetch"]
  }
}

Review

Fetch does one thing: it grabs the text content from any public URL and hands it back to Claude, stripped of HTML noise. No browser automation, no JavaScript rendering, just a straightforward HTTP GET and a clean text extraction. It's the official Anthropic server for this job, which means it's tested against Claude Desktop and works without ceremony. I reach for this when I'm asking Claude to summarise a blog post, pull key points from documentation, or follow a chain of references during research. The workflow is simple: you mention a URL in your prompt, Claude calls the fetch tool, and the text appears in context. No need to copy-paste or switch windows. It's particularly useful for long-form articles where you want Claude to extract specific information without you reading the whole thing first. The limitation is right there in the name: it fetches static HTML. If the content loads via JavaScript, you'll get an empty page or a loading spinner. That rules out most modern SPAs, paywalled sites that require cookies, or anything behind authentication. For those, you need Puppeteer or a headless browser setup. But for public, server-rendered pages, this is faster and lighter than spinning up a browser instance. Setup is trivial if you're on Claude Desktop. One uvx command, a few lines in your config, and it's live. The server has no environment variables to wrangle and no API keys to manage. It just works. I've used it to pull Hacker News threads, GitHub readmes, and API reference pages without friction. The quirk worth knowing: some sites block requests that don't look like a real browser. Fetch doesn't pretend to be Chrome, so you might hit a 403 on aggressive CDNs or bot-protection layers. When that happens, you're back to manual copy-paste. But for the majority of public web pages, this is the no-fuss option. If your workflow involves Claude reading URLs more than once a week, install it.
Verdict

Install this if you regularly ask Claude to read web pages. It's the simplest way to pull public HTML into context without leaving your editor. Skip it if you need JavaScript-rendered content or authenticated pages, where a headless browser is the only real option.

Good at

  • Zero configuration beyond the install command, no API keys or environment variables to manage.
  • Official Anthropic server, so it's maintained alongside Claude and tested for compatibility.
  • Fast and lightweight compared to browser automation for static HTML pages.
  • Works seamlessly in Claude Desktop with automatic tool discovery once configured.
  • Handles the common case of pulling article text or docs without needing a full scraping stack.

Watch out

  • Fails silently on JavaScript-heavy sites that render content client-side, returning empty or incomplete text.
  • Some CDNs and bot-protection services block requests that don't mimic a full browser user agent.
  • No support for authenticated pages or content behind login walls, limiting use to public URLs only.
  • Manual config required for hosts beyond Claude Desktop, though the process is straightforward.

Use cases

  • Pulling article text for summarisation
  • Reading API docs on demand
  • Following links during research
  • Lightweight scraping of static sites

Getting started

1. Run `uvx mcp-server-fetch` in your terminal to confirm the package installs cleanly. 2. Open your Claude Desktop config (usually `~/Library/Application Support/Claude/claude_desktop_config.json` on macOS) and add the server block under `mcpServers`. 3. Restart Claude Desktop, then ask it to fetch a simple public URL like a Wikipedia article to verify the tool appears in the context. 4. Watch out for sites with aggressive bot protection or heavy JavaScript rendering, where you'll get empty or partial text. 5. For other hosts like Cursor or Cline, check their MCP setup docs as the config path and format may differ slightly.

Works with

Claude DesktopClaude CodeCursorWindsurfCline

Similar MCPs