Articles

Google Search Console MCP: A Free Connector for Claude, ChatGPT and Cursor

Our own proposal: before you trust an assistant's summary of your Search Console data, check the order the connector hands it rows in, because Google's API sorts by clicks and breaks ties alphabetically. On a site where most searches get no clicks, a request for the top 25 searches returns the first 25 by letter. On our own property, the search Google showed most often over 28 days came back 16th in the API's default order, behind 15 searches seen once or twice each. So we built a free connector, ModelBrain Search Console, that ranks rows by clicks and then impressions before the assistant reads them.

A row of frosted glass cards in a walnut tray on dark slate, one card raised and glowing amber

What the connector is

ModelBrain Search Console is a hosted MCP server at https://gsc.modelbrain.net/mcp. You add it to Claude, ChatGPT or Cursor, sign in with Google, and your assistant can read every Search Console property your Google account can see, across the full 16 months Search Console keeps. It is free, it is read-only, and you do not need to create a Google Cloud project. The code is open source under the MIT licence on GitHub, and the setup steps for each app are on our connectors page.

Our own proposal: check the order before the summary

Google's Search Analytics API returns rows sorted by clicks, highest first. When rows have the same number of clicks, which on a small or new site is most of them at zero, the API orders them by their text. A connector that asks Google for 25 rows and passes them straight on gives the assistant 25 searches chosen by the alphabet, not by how often people saw your site.

That is what we saw on our own property. In the API's default order for the last 28 days, the first 15 rows were searches seen once or twice each. The search Google had shown most often in that window arrived in 16th place. An assistant asked "what are my top searches?" would have described the first page of the alphabet with confidence.

You can test any Search Console connector the same way. Ask your assistant for your top 10 searches over the last 28 days and look at the list. If it reads like an index, starting with numbers, then words beginning with A, the connector is passing Google's order through. Our connector fetches a wider set of rows and ranks them by clicks, then by impressions, so the top 10 are the 10 searches that mattered most.

How to connect it

In Claude, open Settings, then Connectors, choose Add custom connector, and paste https://gsc.modelbrain.net/mcp. In ChatGPT, turn on developer mode in settings, add a new app or connector with the same address, then pick it from the + menu in a new chat. In Cursor, add the address under mcpServers in your mcp.json. Each app then sends you to Google to sign in.

On Google's permission screen, tick the box that says "View Search Console data for your verified sites". Google leaves it unticked, and without it the connector cannot read anything. Our connectors page has the steps for each app, including the mcp.json block for Cursor.

The five tools

list_sites lists every property the account can see. search_performance returns clicks, impressions, click-through rate and position by search, page, country, device, date or search appearance, with filters. compare_periods puts two date ranges side by side and lists what gained and lost the most. inspect_url reports whether a page is indexed, when Google last crawled it and which canonical it chose. list_sitemaps shows each submitted sitemap with errors and warnings. Every tool is read-only, so the connector cannot submit sitemaps or change settings.

Three more things we fixed while building it

Property names trip assistants up. Search Console calls a domain property sc-domain:example.com, and asking for https://example.com instead returns a permission error even when you own the site. Our connector matches a plain domain or a page address to the right property, and says which one it used.

Comparisons on quiet sites came back empty. When no search changed in clicks between two periods, a click-based comparison has nothing to report, so ours falls back to ranking the changes by impressions and says so.

Sitemap reports carried a number that misleads. Google's sitemap data includes an indexed count that it no longer fills in and that always reads zero. An assistant reading it would tell you none of your pages are indexed, so we leave it out and point to inspect_url for real index status.

What it keeps

The connector stores your Google account identifier, the access token Google issues, encrypted with a key kept outside the database, and its own sign-in tokens for the apps you connect, stored as hashes. Your Search Console data goes straight to the app that asked for it and is never saved. A connection unused for 90 days is deleted. Section 7 of our privacy policy sets this out, including how Google user data is handled.

Why a memory company built a Search Console connector

The connector reads today's numbers. What it cannot do is remember what you and your assistant concluded about them last month. That is the part ModelBrain does: it keeps those notes on your own computer, so the next time you look you can ask what changed since then, in Claude, ChatGPT or Cursor. How ModelBrain works explains the memory side.

Questions people ask

Is there a free Google Search Console MCP server? Yes. ModelBrain Search Console is free for every property your account can see, with no row cap beyond the 25,000 rows Google allows in one report.

Does it work with ChatGPT? Yes, as a custom connector in developer mode. It also works in Claude and Cursor.

Can it request indexing for a page? No. Google does not offer Request Indexing to apps for ordinary pages, so that stays in Search Console. The connector can tell you whether a page is indexed.

Is my Search Console data stored? No. Results pass straight through to your app. Only the sign-in details listed above are kept.