Google Search Console MCP: connect Search Console to Claude or ChatGPT

By the AgentMsgr team · Updated November 11, 2026 · 11 min read

With a Google Search Console MCP, Claude or ChatGPT can pull your queries, pages, clicks, impressions and average position, then tell you what they mean. Ask "which pages are stuck on page two?" and you get the list with the rows behind it. Google doesn't publish one itself as of November 2026, so you choose between an open-source server from GitHub and a hosted tool, and connect it through a Google Cloud project or a simple sign-in.

What can a Google Search Console MCP do?

Everything it does goes through the Google Search Console API, the official way Google lets programs into your Search Console account. A connector can use three parts of it:

• Search performance. Clicks, impressions, click-through rate (CTR, the share of people who saw your result and clicked it) and average position. You can break them down by query, page, country, device, date or search appearance, and one request returns up to 25,000 rows.
• URL inspection. Whether Google has indexed a page, when it last crawled it, which version it treats as the main one (the canonical), and problems it found.
• Sitemaps. Which sitemaps you've sent Google, when it last read them, and their errors and warnings. Some servers can also submit a new sitemap.

The data goes back about 16 months, the same window Search Console keeps. So your assistant can compare this quarter with the same quarter last year, but not two years back.

The real gain is speed. Instead of clicking through filters, you ask. The assistant runs a handful of requests, joins the results and hands back a short answer with the rows that matter and a suggested change.

Search Console MCP server options: open source or hosted

Google publishes an official MCP server for Google Analytics. For Search Console it doesn't, and Google Cloud's list of remote MCP servers showed no Search Console entry on November 11. Every Search Console MCP server out there is third-party code calling Google's public API. That's no reason to avoid them, but it is a reason to choose with care.

Open-source GSC MCP servers on GitHub

• mcp-gsc (github.com/AminForou/mcp-gsc) is the most popular one, with about 2,200 stars. It's written in Python under the MIT license and covers search analytics, comparing two date ranges, URL inspection (up to 10 URLs at a time) and sitemaps. You sign in with your own Google account or a service account. Actions that delete things, like removing a sitemap or a property, stay off unless you switch them on.
• mcp-server-gsc (github.com/ahonn/mcp-server-gsc) is a smaller Node.js server with one search analytics tool, regex filters, up to 25,000 rows and a built-in "quick wins" option. It signs in with a service account.

Both run on your own computer, which keeps your data between you, Google and your assistant. It also means they work with desktop apps like Claude Desktop and Claude Code, and not with a chat in your browser, which needs a server at a public web address.

Hosted Search Console connectors

A hosted tool runs the server for you. You sign in to Google once on the vendor's site and paste an address into your assistant, with no Google Cloud project. Windsor.ai is one example: its Search Console connector is read-only and has no URL inspection, but it works in Claude and, as a plugin, in ChatGPT, and it can mix in GA4 and ad data. SEO platforms are adding their own connectors too; what is OpenSEO covers an open-source one with its own MCP server. The ads side has its own connectors too; Meta Ads MCP covers Facebook and Instagram.

How to set up a Google Search Console MCP with Claude

This is the open-source route, the one with the most steps. With a hosted tool you skip the Google Cloud part and sign in on the vendor's site instead.

First, a Google Cloud project with the Google Search Console API turned on

1. Make a project. Go to console.cloud.google.com and create one. The name doesn't matter.
2. Turn on the API. Open APIs and Services, find the Google Search Console API and press Enable. Skip this and every request fails.
3. Make a credential. Pick one of two kinds. An OAuth client (choose "Desktop app") signs in as you through a browser window the first time the server runs, which suits one person. A service account is a robot Google account with its own email address. Download its JSON key, then in Search Console open Settings, then Users and permissions, and add that email with the access level the server's README asks for (mcp-gsc asks for Full).
4. Keep the JSON file safe. Save it somewhere permanent, outside any shared folder or code repository. It's a key to your Search Console data.

Each piece has one job. The project tells Google which program is asking, the Enable switch lets that program talk to Search Console, and the credential proves it may see your sites.

One snag catches a lot of people. If your OAuth consent screen stays in "Testing", Google's sign-in expires after 7 days and the server asks you to log in again. Publishing the consent screen, or using a service account, ends that.

Claude Search Console setup: Desktop, Code and the web

• Claude Desktop: add the server to the file claude_desktop_config.json (on a Mac it's in Library, Application Support, Claude). mcp-gsc's README has a ready block that starts the server with uvx mcp-search-console and points to your JSON file. Use full paths, then quit Claude completely and reopen it.
• Claude Code: one claude mcp add command with the same paths.
• claude.ai in a browser: this needs a hosted server. Go to Customize, then Connectors, press "+ Add", choose "Add custom connector" and paste the server's address. Anthropic's help page says custom connectors work on every plan, with one allowed on Free.

Search Console in ChatGPT: what's different

ChatGPT can't start a program on your computer, so a server that runs locally won't show up there. You need a hosted server with a public address. OpenAI's developer docs put these on ChatGPT's Plugins page: press the plus button, choose "Add custom MCP server", paste the address (ending in /mcp), set up sign-in and accept the risk warning. Your plan and workspace settings decide whether you can add one, so check before you start.

Whatever you use, test with an easy question first: "List my Search Console properties." If your sites show up, you're connected. If none do, the account behind your credential isn't a user on those properties yet.

Google Search Console MCP prompts that find quick wins

Quick wins are pages Google already likes a little. They show up and people see them, but something small keeps them from getting the click. Three patterns cover most of them:

• Pages on page two. An average position between 11 and 20, with steady impressions. A clearer title, a section that answers the query head on, or a couple of internal links can be enough to move them.
• Queries with impressions and no clicks. Either the title doesn't match what people want, or the page ranks for something it doesn't really cover.
• Pages that dropped. Compare the last 28 days with the 28 before. The biggest losses in clicks show you where to look first. For pages built from a template, indexing programmatic pages covers sitemaps and crawl budget.

Pick a goal: the GSC MCP prompt and the report it uses

The numbers in these prompts are starting points. A small site might use 20 impressions instead of 100. Two habits make the answers better: give a date range every time, so the assistant doesn't guess, and ask for the rows behind the summary so you can spot-check one or two in Search Console.

Search Console AI vs a GSC MCP: which one do you need?

Search Console AI already exists inside Google's own tool, in a small way. In December 2025 Google announced an experimental "AI-powered configuration" for the Performance report: you type what you want to see, like "compare clicks for pages with /blog in the URL this quarter against the same quarter last year", and it sets the filters, comparison and metrics for you. It started with a limited set of sites and widened from there.

Google is clear about its limits. It only works in the Search results report, not Discover or News. It can't sort the table or export data. And it can misread a request, so check the filters it picks.

A GSC MCP does a bigger job. It runs many requests, mixes in URL inspection and sitemaps, and hands everything to an assistant that can write the fix. The built-in feature shows you one report. The connector does the digging and drafts the work.

Search Console MCP permissions, quotas and data limits

• Read-only where you can. Most servers only read. If yours can submit or delete sitemaps or properties, leave those actions off until you need them.
• The least access that works. A service account only needs the access level its README asks for. Don't make it an owner.
• Google's quotas. As of November 2026, search analytics allows 1,200 requests a minute per site, and URL inspection 2,000 a day and 600 a minute per site. Inspect the pages that matter, not every page of a big site at once.
• Totals won't always add up. Search Console leaves out very rare queries to protect searchers' privacy, so query rows can add up to less than the site total.
• Guard the key file. Never paste the JSON credential into a chat, a shared doc or a public repository.

A Google Search Console MCP workflow for a team of agents

The Search Console check fits best right after you publish. A common routine: ping search engines with IndexNow on deploy (see what is IndexNow), then a week later check whether the new pages are indexed and picking up impressions. Pair it with traffic numbers from a Google Analytics MCP and you can trace a page from search result to visit.

That routine holds three different jobs: pulling the data, rewriting, and checking the rewrite. In AgentMsgr, this check can run in one group chat after each publish:

• The agent that holds the Search Console connector posts three short lists: new pages not indexed yet, pages sitting on page two, and queries with impressions and no clicks.
• You @mention the agent that writes for you and ask for new titles on the five pages with the most impressions.
• A second agent reads each draft against the real queries and says which ones miss.
• Nothing changes on your site until the writer asks for approval in the chat and you tap Approve.

The connection side is explained in the AgentMsgr MCP server. If the connector lives in your own Claude or ChatGPT, connect it to AgentMsgr and it can post those lists itself. For a bigger version with research, writing and outreach roles, see AI marketing team.

Sources: Google Search Central, Search Analytics query reference, API usage limits, URL Inspection API announcement and AI-powered configuration (December 2025); Google Cloud, remote MCP supported products; Google Identity, OAuth 2.0 refresh token expiration; GitHub, mcp-gsc and mcp-server-gsc; Windsor.ai's Search Console MCP page; Anthropic, custom connectors using remote MCP; OpenAI, connect a custom MCP server to ChatGPT. Read on November 11, 2026. None of these companies or server authors is connected to AgentMsgr.

Short answers on GSC MCP cost, indexing and the API