Fetch MCP Server

@yawlabs/fetch-mcp

One click adds this to your local Yaw MCP config so it's available in every Yaw Terminal session. Or install manually below.

A comprehensive HTTP fetch MCP server for AI assistants. Bring-your-own client: runs as a stdio MCP server so any MCP-compatible client (Claude Code, Claude Desktop, Cursor, mcph, …) can fetch web content safely.

What it gives the model

Safety

SSRF protection is on by default. The server refuses requests to:

• Loopback (127.0.0.0/8, ::1)
• RFC1918 private ranges (10/8, 172.16/12, 192.168/16)
• Link-local (169.254/16, fe80::/10) — including the cloud metadata endpoint 169.254.169.254
• CGNAT (100.64/10)
• Unique-local IPv6 (fc00::/7)
• Multicast / broadcast
• IPv4-mapped IPv6 (::ffff:0:0/96) re-checked against the IPv4 rules, and likewise the IPv4 embedded in IPv4-compatible (::/96), IPv4-translated (::ffff:0:0:0/96) and 6to4 (2002::/16) addresses
• Teredo (2001::/32), NAT64 (64:ff9b::/96, 64:ff9b:1::/48), site-local (fec0::/10) and discard-only (100::/64) IPv6
• Any other IPv6 outside global unicast (2000::/3), and the non-global blocks inside it (3fff::/20, 2001:2::/48, 2001:10::/28)
• The Azure WireServer address 168.63.129.16
• Non-http/https schemes (file://, gopher://, javascript:, …)
• Hostname localhost and any *.localhost

Every redirect hop is re-validated against all of the above -- scheme, literal IP and localhost* -- before it is dialed, so a 302 from a public host to http://127.0.0.1, http://169.254.169.254 or ftp://… is caught. For hostnames, DNS is resolved once per hop, every returned address is checked, and the verified IP is pinned into the HTTP dispatcher so the subsequent TCP connection dials that exact address — closing the DNS-rebinding TOCTOU window. Authorization, Cookie and Proxy-Authorization headers are stripped on cross-origin redirects.

Reaching private hosts (development)

The model chooses tool arguments, and the model is not trusted: a prompt injected through a fetched page can ask for anything. So the per-call allow_private_hosts: true opt-in is refused unless the operator enables it by launching the server with FETCH_MCP_ALLOW_PRIVATE_HOSTS=1 (true/yes/on also work; any unrecognised value is treated as off and named on stderr). With the variable set, a call still has to ask: requests without allow_private_hosts stay guarded.

{
  "mcpServers": {
    "fetch": {
      "command": "npx",
      "args": ["-y", "@yawlabs/fetch-mcp@latest"],
      "env": { "FETCH_MCP_ALLOW_PRIVATE_HOSTS": "1" }
    }
  }
}

Only set it where the model may legitimately reach your internal network -- a local dev box, not a cloud VM with a metadata endpoint.

Install & run

# One-off (auto-updates each spawn via @latest)
npx -y @yawlabs/fetch-mcp@latest

# Or globally
npm i -g @yawlabs/fetch-mcp
fetch-mcp

Requires Node 22.19 or newer when running on Node (the floor of its HTTP client, undici 8; the launcher refuses older Nodes with a clear message instead of crashing). Under oam, which the launcher prefers when installed, the Node version does not matter.

Configure in Claude Code / Claude Desktop

Add to your client's MCP config (usually claude_desktop_config.json or ~/.claude.json):

{
  "mcpServers": {
    "fetch": {
      "command": "npx",
      "args": ["-y", "@yawlabs/fetch-mcp@latest"]
    }
  }
}

Or via mcph:

mcph add fetch

Tool reference

http_get, http_post, http_put, http_patch, http_delete, http_head, http_options

Common parameters:

Body-capable tools (POST/PUT/PATCH/DELETE) also take:

Response shape:

{
  ok: boolean;
  status: number;
  statusText: string;
  url: string;              // final URL after redirects
  headers: Record<string,string>;
  body_text?: string;
  body_base64?: string;     // when decode_text=false
  json?: unknown;           // auto-parsed when response is application/json
  truncated?: boolean;      // set when max_bytes hit
  redirects?: string[];     // chain of intermediate URLs
  duration_ms: number;
  error?: string;
}

fetch_html_to_markdown

GET the URL, strip scripts/styles/iframes/svg/canvas plus <nav>, <footer>, <aside>, convert to atx-headed markdown with fenced code blocks and dash bullets. Intended for feeding pages into an LLM without blowing the context budget.

fetch_html_to_text

Same fetch, but emits plain text with block-level structure preserved as newlines. Useful when the model doesn't need markdown formatting.

fetch_reader

Isolates the main article body using, in order: <article>, <main>, itemprop="articleBody", common CMS class names (post-content, entry-content, etc.), then <body> as fallback. Returns:

{
  url: string;          // final URL after redirects
  title?: string;       // og:title, then <title>, then <h1>
  byline?: string;      // meta[name=author] / article:author
  wordCount: number;
  markdown: string;     // main content converted to markdown
}

fetch_meta

GET a URL and return its head metadata without downloading the full body (caps at 2 MiB by default):

{
  url: string;
  title?: string;
  description?: string;
  canonical?: string;
  language?: string;
  robots?: string;
  og:      Record<string, string>;       // first value per key (og:title, og:image, og:type, ...)
  twitter: Record<string, string>;       // first value per key
  article: Record<string, string>;       // first value per key
  ogAll:      Record<string, string[]>;  // only keys that appear > once (e.g. multiple og:image)
  twitterAll: Record<string, string[]>;
  articleAll: Record<string, string[]>;
  icons:   Array<{ rel: string; href: string; sizes?: string }>;
  feeds:   Array<{ href: string; title?: string; type?: string }>;   // RSS/Atom
  jsonLd:  unknown[];                    // parsed application/ld+json blocks
}

fetch_links

GET a page and return every <a href> with text, resolved to absolute URLs. Respects <base href>. Skips #, javascript:, mailto:, tel:, data:, file:. Each link is classified internal or external vs. the page host. Optional filter/dedupe/limit.

fetch_sitemap

Fetch a sitemap.xml or sitemap-index and return the URL list:

{
  sitemaps: string[];       // indexes followed, in order
  urlCount: number;
  truncated: boolean;       // hit max_urls
  childSitemaps: string[];  // listed but not fetched (past max_depth or max_sitemaps)
  warnings: Array<{ url: string; error: string }>;
  urls: Array<{
    loc: string;
    lastmod?: string;
    changefreq?: string;
    priority?: number;
  }>;
}

Gzipped .xml.gz sitemaps are auto-decompressed. max_depth controls how many levels of sitemap-index to follow (default 1). Setting max_depth: 0 on a sitemap-index returns the index's childSitemaps list without fetching any child (useful to discover structure cheaply). Partial failures — one child sitemap 500s while others succeed — are returned under warnings rather than aborting the whole call. max_sitemaps (default 50, max 1000) caps how many sitemap documents one call fetches, index included; children past it are listed under childSitemaps with a warning. The whole call is capped at 5 minutes. max_bytes (default 20 MiB) applies to each sitemap, and to the decompressed size of a gzipped one; a sitemap larger than that is an error (a warning for a child) rather than a silently partial URL list.

fetch_feed

Parse an RSS 2.0 or Atom 1.0 feed:

{
  kind: "rss" | "atom" | "unknown";
  title?: string;
  description?: string;
  link?: string;
  updated?: string;
  entryCount: number;
  truncated: boolean;       // hit limit
  entries: Array<{
    title?: string;
    link?: string;
    id?: string;
    published?: string;
    updated?: string;
    author?: string;
    summary?: string;
    content?: string;
    categories?: string[];
  }>;
}

fetch_robots

Fetches <origin>/robots.txt, parses it, and returns:

{
  robotsUrl: string;
  status: number;
  userAgent: string;
  path: string;
  allowed: boolean;
  matchedRule: string | null;  // the longest-match Allow/Disallow that decided it
  crawlDelay: number | null;   // from the matched group
  sitemaps: string[];          // top-level Sitemap: declarations
  rawRobotsText: string;       // first 512KB
  note?: string;               // only when robots.txt 404s
}

A /robots.txt that returns 404 means there are no rules, so the verdict is allowed: true. The response keeps the same shape: status: 404, matchedRule: null, crawlDelay: null, sitemaps: [], rawRobotsText: "", plus the note. Any other non-2xx status is returned as an error.

The parser follows Google's rules: longest match wins, * is a wildcard segment, $ anchors the end of the path, specific user-agent group beats the * wildcard group when the UA matches (comparison uses the length of the actually-matched agent token, not the group's first agent). Allow beats Disallow on equal-length ties.

Development

npm install
npm run build
npm test         # vitest
npm run lint     # biome
npm run typecheck

Tests spin up a local loopback HTTP server on 127.0.0.1:0 to exercise the real request/response path — no mocking of HTTP. SSRF tests verify that the default-deny still applies to that local server unless the operator gate is on and the request opts into allow_private_hosts.

License

MIT © Yaw Labs

Links

• npm: https://www.npmjs.com/package/@yawlabs/fetch-mcp
• issues: https://github.com/YawLabs/fetch-mcp/issues
• Yaw Labs: https://yaw.sh