Google search MCP server for Claude Desktop, Cursor and other MCP clients

What does the server do?

It exposes one tool, google_search, over the Model Context Protocol. When the model calls it, the server sends one request to https://serpcrypto.com/api/v1/search, then turns the JSON into short plain text: the AI Overview if Google showed one, and the top results with title, link and snippet.

Short text matters. A full results page in JSON is long, and every character sent back to the model uses context. The server keeps up to ten results and cuts the overview at 800 characters; change those limits in the code if your client has room for more.

This is copy-paste code, not a published package: there is no official SerpCrypto MCP package to install. You own the file and can add tools, for example one for engine=google_news.

"""serpcrypto_mcp.py: an MCP server that gives AI clients a Google search tool.

Install once:  pip install "mcp[cli]" httpx
Run by your MCP client over stdio; the API key comes from the environment.
"""
import os

import httpx
from mcp.server.fastmcp import FastMCP

API_URL = "https://serpcrypto.com/api/v1/search"
mcp = FastMCP("serpcrypto-search")


@mcp.tool()
async def google_search(query: str, gl: str = "us", hl: str = "en", num_results: int = 5) -> str:
    """Search Google and return the top results (title, link, snippet) and the AI Overview text if any.

    gl is a two-letter country code (us, gb, de...), hl a language code (en, de, pt-BR...).
    Each call uses one SerpCrypto search credit.
    """
    headers = {"Authorization": "Bearer " + os.environ["SERPCRYPTO_API_KEY"]}
    params = {"q": query, "gl": gl, "hl": hl}
    async with httpx.AsyncClient(timeout=90) as client:
        r = await client.get(API_URL, params=params, headers=headers)
    if r.status_code != 200:
        return f"Search failed with HTTP {r.status_code}: {r.text[:300]}"
    data = r.json()
    lines = []
    overview = data.get("ai_overview") or {}
    if overview.get("text"):
        lines.append("AI Overview: " + overview["text"][:800])
    for item in data.get("organic_results", [])[: max(1, min(num_results, 10))]:
        lines.append(f"{item['position']}. {item['title']}\n{item['link']}\n{item.get('snippet', '')}")
    return "\n\n".join(lines) or "No results."


if __name__ == "__main__":
    mcp.run()  # stdio transport

Why is the tool called google_search and not web_search?

Several clients already ship a tool named web_search that uses their own provider. A distinct name avoids a clash and tells the model what it gets: Google results for a country and language it chooses. The docstring is the description the model reads, so it states the parameters and the cost of a call.

How do I add it to Claude Desktop?

Open the configuration file, claude_desktop_config.json: on macOS it is in ~/Library/Application Support/Claude/, on Windows in %APPDATA%\Claude\. Add the server under mcpServers, with the absolute path to the script and your key in env, then restart Claude Desktop.

If python is not on the path Claude Desktop sees, use the full path to the interpreter, for example the one inside your virtual environment.

{
  "mcpServers": {
    "serpcrypto-search": {
      "command": "python",
      "args": ["/absolute/path/to/serpcrypto_mcp.py"],
      "env": {
        "SERPCRYPTO_API_KEY": "paste-your-key-here"
      }
    }
  }
}

How do I add it to Claude Code and Cursor?

In Claude Code, register it from a terminal with claude mcp add; -e sets the environment variable and everything after -- is the command to run. In Cursor, put the same mcpServers block as above in ~/.cursor/mcp.json for every project, or in .cursor/mcp.json inside one project.

Other MCP clients that start stdio servers take the same three things: a command, its arguments and environment variables. The OpenClaw page shows the commands for OpenClaw.

claude mcp add serpcrypto-search \
  -e SERPCRYPTO_API_KEY=paste-your-key-here \
  -- python /absolute/path/to/serpcrypto_mcp.py

How do I keep the API key safe?

Keep it in the env block of the client configuration or in your shell profile, never inside serpcrypto_mcp.py. That way the script can be shared or committed without the key. If a key leaks, revoke it in the dashboard and create a new one: API keys you can create and revoke at any time from your dashboard.

Give the MCP server its own key. Usage per key then shows how many searches the assistant makes, separate from your scripts.

How many credits will an assistant use?

One per tool call that succeeds. A chat where the model searches three times costs three credits. On the 25,000 searches pack at $18.75, that is thousands of assisted questions, and Prepaid search credits never expire: unused searches stay on your account until you use them.

To cap usage, lower num_results (it does not change the price, but it shortens answers) and tell the model in your instructions to search only when it needs fresh facts.