Google Search Console MCP Server: Connect It to Claude Code

mcp-gsc is a Python MCP server that connects Google Search Console to Claude Code. Once connected, Claude can pull search analytics, inspect URLs, check indexing status, and manage sitemaps — all through tool calls.

Update: mcp-gsc now runs with uvx

Since April 2026, mcp-gsc has been on PyPI as mcp-search-console, and its README now recommends running it with uvx instead of cloning the repo.

Set "command" to the full path of uvx, set "args" to ["mcp-search-console"], and pass the path to your client_secrets.json in the GSC_OAUTH_CLIENT_SECRETS_FILE environment variable:

1
{
2
  "mcpServers": {
3
    "gsc": {
4
      "command": "/full/path/to/uvx",
5
      "args": ["mcp-search-console"],
6
      "env": {
7
        "GSC_OAUTH_CLIENT_SECRETS_FILE": "/path/to/client_secrets.json"
8
      }
9
    }
10
  }
11
}

The clone steps below still work on Python 3.11 or newer, and the README now lists 20 tools.

Corrected: when an allow and a disallow rule match the same path, Google uses the least restrictive one, so the robots.ts rules described below never blocked Googlebot (Google’s robots.txt spec).

I set it up for 0xinsider.com and ran an audit. The first thing it found was that my robots.ts file had disallow: ["/"] sitting next to allow: "/", two rules that contradict each other. That had been live for weeks.

Without this, here’s how that would’ve gone: open Search Console, notice low impressions, scratch my head, check a few pages manually, eventually think to look at robots.txt, open VS Code, find the file, spot the conflict, fix it. Thirty minutes if I’m lucky, days if I don’t think to check robots.

With the MCP server, Claude pulled the performance data and the robots config in the same turn. It saw low impressions next to a misconfigured crawl rule and connected the dots. I typed “fix it” and the file was updated. The whole thing took seconds.

That’s the real value — not just reading data faster, but closing the loop between diagnosis and fix. Every SEO workflow I’ve done before involved jumping between the Search Console dashboard, a spreadsheet, and my editor. Now it’s one conversation.

Setup

The server uses OAuth to authenticate with your Google account. You need a Google Cloud project with the Search Console API enabled.

1. Go to Google Cloud Console, create a project (or use an existing one)
2. Enable the Search Console API
3. Go to Credentials, create an OAuth 2.0 client ID (Desktop app type)
4. Download the JSON file as client_secrets.json

Clone the server and install dependencies:

1
git clone https://github.com/AminForou/mcp-gsc.git
2
cd mcp-gsc
3
uv venv .venv && source .venv/bin/activate
4
uv pip install -r requirements.txt

Place your client_secrets.json in the mcp-gsc directory. On first run, the server opens a browser for OAuth consent. The token gets saved to token.json.

Add it to your Claude Code MCP config:

1
{
2
  "mcpServers": {
3
    "gsc": {
4
      "command": "/path/to/mcp-gsc/.venv/bin/python",
5
      "args": ["/path/to/mcp-gsc/gsc_server.py"]
6
    }
7
  }
8
}

Restart Claude Code. You should see GSC tools available.

What it found on 0xinsider.com

I asked Claude to investigate search performance. It pulled performance overview, top queries, top pages, device breakdown, country data, sitemaps, and URL inspection results — all in parallel, in one turn.

28-day numbers:

Here’s what it found:

robots.ts had contradicting rules. The file had both allow: "/" and disallow: ["/"]. Google picks the rule with the longest matching path, and when two rules tie, as these do, it uses the least restrictive one, so Google could still crawl everything. The disallow did nothing, so I scoped it to /api/. Here’s what it looked like:

1
// Before (broken)
2
rules: {
3
  userAgent: "*",
4
  allow: "/",
5
  disallow: ["/"],
6
}
7

8
// After (fixed)
9
rules: {
10
  userAgent: "*",
11
  allow: "/",
12
  disallow: ["/api/"],
13
}

The learn guide had a search intent mismatch. The page /learn/how-to-trade-prediction-markets had 1,411 impressions and 1 click — a 0.07% CTR. The queries driving those impressions were about “how to invest in prediction markets through stocks,” not about trading on Polymarket directly. The title said “How to Trade” but users were searching for “How to Invest.” Changed the title to match the actual queries.

Leaderboard wasn’t ranking for its own keyword. The /leaderboard page was showing up at position 12-48 for “polymarket leaderboard” queries. The title was “Top Polymarket Traders — Most Advanced Trader Rankings.” It didn’t contain the word “leaderboard.” Changed it to “Polymarket Leaderboard — Top Traders Ranked by Profit.”

Trader profiles had generic meta descriptions. Every profile page had the same template: “Deep dive into @name’s Polymarket strategy. Full P&L breakdown…” — no actual stats. Updated generateMetadata to fetch trader data and include real numbers: P&L, win rate, markets traded. So now a profile shows something like “@gabagool22 Polymarket Profile — $214K P&L, 52% win rate” in search results instead of a generic blurb.

Desktop CTR was 5x worse than mobile. Desktop: 4,321 impressions, 0.74% CTR. Mobile: 767 impressions, 3.52% CTR. The meta descriptions were too long and generic for desktop SERPs where Google shows more text. Shortened them and added specific numbers.

All seven fixes happened in the same Claude Code session where the data was pulled. No tab switching, no exporting CSVs, no copying query data into a spreadsheet.

The tools

The server exposes 19 tools. The ones I used most:

batch_url_inspection is the one that saves the most time. In the Search Console UI, you inspect URLs one at a time. With the MCP tool, you pass 5 or 10 URLs and get back indexing status and rich result detection for all of them in one call.

get_search_by_page_query is useful for diagnosing CTR problems. You see a page with 30,000 impressions and 0.11% CTR, so you pull the queries for that page and discover the title doesn’t match what people are actually searching for.

What makes this different from just using the GSC dashboard

The dashboard shows data. Claude Code shows data and has access to the codebase at the same time.

When Claude finds that a page has high impressions and low CTR, it can read the page’s frontmatter, see the current title and meta description, compare them against the actual search queries, and propose a change — all without you navigating anywhere. The gap between finding a problem and fixing it is one sentence: “fix it.”

The robots.txt bug is a good example. The performance data alone didn’t point at it. The robots data alone looked like a normal config file.

It was the combination, seeing poor performance alongside a misconfigured robots file, that got it flagged and fixed.