Google Search Web Fetch MCP

Dockerized MCP server exposing two read-only tools:

• google_search: searches Google through a CloakBrowser stealth Chromium session and returns parsed organic results.
• web_fetch: renders a page through CloakBrowser and returns article, body, selector, or raw content as markdown, text, or HTML.

The HTTP transport follows the current MCP Streamable HTTP convention with a single MCP endpoint at /mcp. Stdio is also available for local MCP clients that spawn servers as subprocesses.

Build

docker compose build

The image build installs Node dependencies, browser system dependencies, and fonts. On container startup, the server warms CloakBrowser before accepting HTTP requests, so Chromium install/update work happens during startup instead of the first tool call. No local Node install is required.

Run HTTP

docker compose up

Endpoints:

• POST /mcp: MCP JSON-RPC requests.
• GET /mcp: MCP SSE stream for Streamable HTTP sessions.
• DELETE /mcp: MCP session termination.
• GET /healthz: container health.
• GET /readyz: readiness and active session count.

Default URL:

http://127.0.0.1:3000/mcp

Run Stdio

Build the image first, then configure a local MCP client to spawn the container:

{
  "mcpServers": {
    "google-search-web-fetch": {
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "-e",
        "MCP_TRANSPORT=stdio",
        "google-search-web-fetch-mcp:latest"
      ]
    }
  }
}

Tool Inputs

google_search:

{
  "query": "model context protocol streamable http",
  "count": 10,
  "start": 0,
  "language": "en",
  "region": "US",
  "safe": "active",
  "timeoutMs": 30000,
  "stealth": true,
  "humanize": false,
  "fingerprint": "42069"
}

web_fetch:

{
  "url": "https://modelcontextprotocol.io/specification/2025-11-25/basic/transports",
  "format": "markdown",
  "extract": "article",
  "waitUntil": "domcontentloaded",
  "timeoutMs": 30000,
  "maxContentChars": 50000,
  "includeLinks": false,
  "stealth": true,
  "humanize": false
}

Both tools return MCP content text plus structuredContent matching their advertised output schema.

Configuration

Copy .env.example if you want local overrides. Common settings:

• MCP_PORT: HTTP port, default 3000.
• MCP_ALLOWED_ORIGINS: comma-separated allowed browser origins. Defaults to localhost origins.
• MCP_BEARER_TOKEN: optional bearer token required on /mcp.
• MCP_ENABLE_FETCH: default true; set false to disable and hide the web_fetch tool.
• MCP_ALLOW_PRIVATE_NETWORK: default false; blocks loopback/private network fetch targets for SSRF protection.
• MCP_BROWSER_WARMUP: default true; launch CloakBrowser during HTTP startup to avoid first-request cold-start latency.
• MCP_BROWSER_UPDATE_ON_START: default true; check for and download newer CloakBrowser Chromium during startup warmup.
• CLOAK_HEADLESS: default true.
• CLOAK_HUMANIZE: default false.
• CLOAKBROWSER_AUTO_UPDATE: set to false in the Docker image/compose file so CloakBrowser does not start a background update during the first request; startup warmup handles the update check explicitly.
• CLOAK_PROXY: optional proxy URL.
• CLOAK_FINGERPRINT: optional stable fingerprint seed.
• CLOAK_EXTRA_ARGS: comma-separated extra Chromium flags.

Acknowledgements

This project builds on open-source browser automation, MCP, parsing, and TypeScript tooling. See THIRD_PARTY_NOTICES.md for direct dependency acknowledgements and license references.

Notes

CloakBrowser stealth is enabled by default. The server does not solve CAPTCHA challenges; if Google returns an automated-query interstitial, google_search returns a tool error with remediation hints.

Relevant references:

• MCP Streamable HTTP transport: https://modelcontextprotocol.io/specification/2025-11-25/basic/transports
• MCP tool result structured content: https://modelcontextprotocol.io/specification/2025-11-25/server/tools
• CloakBrowser Docker and JavaScript API: https://github.com/CloakHQ/CloakBrowser