fetchmcp

fetchmcp

A drop-in replacement for the official fetch MCP that actually works on modern web pages.

The official fetch MCP is broken on JavaScript-heavy pages, truncates output at 5,000 characters, and ships an unpatched SSRF vulnerability. fetchmcp returns clean, LLM-ready Markdown from any URL — rendering JavaScript when needed, passing basic bot protection without paid proxies, and telling you honestly when a page is blocked instead of hallucinating content. npx and go.

// Replace the official fetch server with this — one line in your MCP config:
"fetchmcp": { "command": "npx", "args": ["-y", "@labtoolsstudio/fetchmcp"] }

 

Why switch

	official fetch	fetchmcp
JavaScript pages	❌ empty / broken	✅ auto-renders in a real browser
Output length	✂️ truncated at 5,000 chars	✅ full page, with paging
Bot protection (403 / Cloudflare)	❌ fails silently	✅ passes mid-tier walls, no paid proxy
Blocked page	❌ returns the CAPTCHA as "content"	✅ honest typed error, never fakes it
SSRF safety	❌ CVE-2025-65513 (CVSS 9.3)	✅ private/metadata IPs refused by default
Cost	free	free, self-hosted, $0

Install

Add to your MCP client config (claude_desktop_config.json, Cursor mcp.json, Cline, etc.):

{
  "mcpServers": {
    "fetchmcp": {
      "command": "npx",
      "args": ["-y", "@labtoolsstudio/fetchmcp"]
    }
  }
}

The install is light — no browser is downloaded up front, and static reading (fetch → Readability → Markdown) works immediately. The first time a page actually needs JavaScript, fetchmcp downloads a stealth Chromium once (~150 MB) automatically, then renders it — still zero-config. To pre-download it at install time, set FETCHMCP_PREINSTALL_BROWSER=1. To stay static-only and never download it, set FETCHMCP_SKIP_BROWSER_DOWNLOAD=1 (JS pages then return an honest needs_js).

Tools

read_url

Fetch any web page as clean Markdown.

arg	type	description
url	string	the URL to fetch (http/https)
render	boolean	JS rendering: true = always, false = never, omitted = automatic (only for empty SPA shells)
raw	boolean	return raw HTML instead of Markdown
headers	object	extra request headers, e.g. {"Authorization": "Bearer …", "Cookie": "…"}
max_length	integer	cap characters returned (0 = unlimited, the default)
start_index	integer	offset for paging through a long page

read_docs

Same engine, tuned for documentation: strips navigation sidebars, headers, and footers so API docs and guides come back as clean reference text. Takes url, render, headers, max_length, start_index.

Honest statuses

fetchmcp never returns a bot wall, an error page, or a truncated shell dressed up as real content. When it can't read a page it says why, with a typed status: blocked (bot protection, with the vendor), blocked_ssrf, needs_js, http_error, timeout, network_error, unsupported_content, or empty.

Configuration (env vars)

var	default	meaning
FETCHMCP_TIMEOUT_MS	30000	per-request timeout
FETCHMCP_MAX_RETRIES	2	retries on network errors / 429 / 503 (with backoff + Retry-After)
FETCHMCP_ALLOW_PRIVATE_IP	unset	set to 1 to allow private/localhost IPs (trusted intranet docs)
FETCHMCP_FLARESOLVERR_URL	unset	self-hosted FlareSolverr endpoint for tougher challenges
FETCHMCP_SKIP_BROWSER_DOWNLOAD	unset	set to 1 for static-only: never download Chromium; JS pages return needs_js
FETCHMCP_PREINSTALL_BROWSER	unset	set to 1 to download Chromium at install time instead of on first JS use

Development & testing

npm install          # installs deps (Chromium downloads on first JS use)
npm run build        # compile TypeScript to dist/
npm test             # unit tests (block detection, SSRF) — no network
npm run test:e2e     # live end-to-end suite against real sites

# Poke at any tool/URL by hand — no need to write a script:
node test/probe.mjs read_url  https://example.com
node test/probe.mjs read_url  https://some-spa.example.com --render
node test/probe.mjs read_docs https://docs.python.org/3/library/json.html
node test/probe.mjs read_url  https://api.example.com --header "Authorization=Bearer x" --max-length 500
node test/probe.mjs read_url  https://example.com --full     # print the whole response

test/probe.mjs --help semantics are documented at the top of that file.

How it works

Three tiers, escalating only as needed:

1. Static — plain fetch → Readability → Markdown. Fast path for most pages.
2. Browser — lazy patchright (stealth Chromium) when the static HTML is an empty SPA shell, a bot wall, or a 403/429/503.
3. FlareSolverr (optional) — only if you've configured an endpoint, for challenges the browser can't clear.

Star history

If fetchmcp saved you from one more fetch-returns-nothing moment, a star helps others find it.

License

MIT — see LICENSE.