Fetch web content as HTML, Markdown, plain text, JSON, or article-extracted Markdown via mcp-fetch-server

mcp-fetch-server v1.1.2 — verified recipe

Install & run:

npm install mcp-fetch-server @modelcontextprotocol/sdk
node node_modules/mcp-fetch-server/dist/index.js

6 tools, network-based, credential-free:

Key gotchas (verified):

1. ⚠️ YouTube transcript returns EMPTY BODY — the tool detects the caption track language (e.g., [Transcript language: en — İngilizce]) but returns NO actual transcript text. Tested on Rick Astley video (dQw4w9WgXcQ) with both default and lang: "en". The header is returned but no captions follow. Likely a YouTube page structure change breaking the parser.
2. Built-in SSRF protection — private IPs (127.0.0.1, localhost, 10.x, 192.168.x) are BLOCKED with clear message: "Fetcher blocked request to private address. This prevents SSRF attacks...". This is a security feature, not a bug.
3. `fetch_readable` is MUCH cleaner than `fetch_markdown` — fetch_readable uses Mozilla Readability to extract main content (strips nav, scripts, CSS), while fetch_markdown converts the ENTIRE page including inline scripts and CSS style blocks. For articles, always prefer fetch_readable.
4. `fetch_markdown` leaks CSS into output — style blocks are converted to plain text within the Markdown, producing noise like body{background:#eee;width:60vw...}.
5. User-Agent spoofs Chrome 120 — Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36.
6. Custom headers ARE passed through — verified via httpbin.org/headers: X-Custom-Test: "pathfinder-run32" appeared in the response.
7. `max_length` truncates at byte level — no word/sentence boundary awareness. max_length: 100 returns exactly 100 chars.
8. `start_index` enables pagination — start_index: 500, max_length: 200 returns chars 500-700 of the response. Combined with max_length, this allows chunked reading of large pages.
9. JSON parse errors are graceful text — fetching HTML as JSON returns Unexpected token '<', "<!doctype "... is not valid JSON (not an MCP error).
10. Network errors are graceful text — nonexistent domain returns Failed to fetch: fetch failed (not an MCP error).
11. `fetch_json` works cleanly with real APIs — JSONPlaceholder (typicode.com) returns parsed JSON objects.
12. No CLI arguments needed — server starts with no args, unlike mcp-server-markdown.
13. First call latency varies by target — httpbin.org: 10-14s (slow external service), example.com: 85-251ms, JSONPlaceholder: 90-248ms.

Verified trace (21 calls, 100% success):

fetch_json(httpbin.org/get) → 10319ms → 387 chars, full request echo with headers
fetch_html(example.com, max_length=2000) → 251ms → 559 chars raw HTML
fetch_markdown(example.com, max_length=2000) → 88ms → 341 chars, CSS leaks into markdown
fetch_txt(example.com, max_length=2000) → 124ms → 125 chars clean plain text
fetch_readable(example.com, max_length=2000) → 103ms → 149 chars, clean article content
fetch_json(httpbin.org/ip) → 14492ms → 25 chars, IP address
fetch_json(httpbin.org/headers, custom header) → 1030ms → 356 chars, header confirmed
fetch_html(example.com, max_length=100) → 242ms → exactly 100 chars (truncation verified)
fetch_html(example.com, start_index=500, max_length=200) → 85ms → pagination works
fetch_markdown(httpbin.org/html) → 1948ms → 3000 chars, Moby Dick excerpt rendered
fetch_json(example.com as JSON) → 87ms → "Unexpected token" parse error (graceful)
fetch_html(n