google-search-mcp

MCP server that gives AI agents Google search and web scraping through a stealth browser (super-puppeteer). Runs headful by default — Google serves a captcha wall to headless Chrome on datacenter IPs.

Tools

| Tool | What it does | |------|--------------| | google_search | Organic results (max 20): title, url, snippet. fetchContent: N also fetches the top N result pages as compressed markdown. Pagination, lang, region. | | google_images | Full-size image URL, dimensions, source page, title, thumbnail. | | google_news | News tab: headline, source, snippet, url. | | scrape_page | Any URL → boilerplate-free markdown. Optional query keeps only the most relevant blocks within maxChars. | | extract_images | All images from any page (img, og:image) with alt + dimensions. | | screenshot | PNG screenshot returned as image content the agent can see. |

Token economy

Fetched page content goes through two passes so agents don't burn tokens on boilerplate:

1. Readability extraction (in-browser) — strips nav/header/footer/ads/cookie banners, picks the densest content container, drops blocks with >60% link density (menus, link farms), emits ordered markdown blocks (headings, paragraphs, lists, quotes, code, tables).
2. Relevance budget (server-side) — if content exceeds the character budget, blocks are scored by query-term frequency density (TF / sqrt(len)) plus heading and early-position bonuses; the highest-signal blocks are kept in original order with … marking gaps.

Requirements

• Node.js >= 23 (runs TypeScript natively)
• Chrome (installed automatically by puppeteer)

Usage (npx, no install)

npx -y @fdcicyber/google-search-mcp

Claude Code / MCP client config

{
  "mcpServers": {
    "google-search": {
      "command": "npx",
      "args": ["-y", "@fdcicyber/google-search-mcp"]
    }
  }
}

Or with Claude Code CLI:

claude mcp add google-search -- npx -y @fdcicyber/google-search-mcp

First run downloads Chrome via puppeteer (cached in ~/.cache/puppeteer, one-time).

Notes

• Browser launches lazily on first tool call and is reused across calls; one page per request.
• Google consent cookies are pre-seeded so the GDPR interstitial never appears.
• Keep it headful. Headless Chrome from this network gets google.com/sorry (unusual traffic captcha).