Google Search Console MCP Server

MCP (the Model Context Protocol) is the standard way for an AI assistant to call tools in another service. Copilot SEO runs an MCP server, so Claude Code, Claude.ai and Claude Desktop can read your Search Console data and, if you allow it, add annotations and start experiments. Other MCP clients can connect with a personal access token.

What can you ask it?

"Which queries does /pricing rank on page two for?"
"What's at the top of the Action plan for this site?"
"Did clicks drop after last Tuesday's release?"
"Mark today's deploy on the chart as v1.4.2."
"Start an experiment for the title change I shipped this morning."
"How is the experiment on the category pages doing?"

The assistant picks the tools and arguments. The figures are the same ones the Overview, Pages and Queries screens show, so they match what you see in the app.

The tools

Tool	Writes	What it returns or does
`whoami`		The person, their teams and roles, and whether the token may write
`setup_status`		Each team's setup steps, done or not, what to do next, and which need the person, with the link
`list_search_console_sites`		Search Console sites the person's Google accounts can see, marking ones already added
`add_site`	Yes	Add a Search Console site to a team and start its import
`set_brand_terms`	Yes	Replace a team's brand terms (owners and admins)
`list_properties`		Sites with their ids and `data_through_date`
`sites_overview`		Every site side by side for a period against the one before, with a trend and actions in Now
`gsc_metrics`		Clicks, impressions, CTR and position by query, page or date
`get_rankings`		How many queries or pages rank in the top 3, the rest of page one, page two and beyond, each week, and what moved between them
`list_recommendations`		The Action plan, highest value first
`get_recommendation`		One action: diagnosis, evidence, how to act, status, review date
`update_recommendation`	Yes	Move an action between Now, Next and Later, or dismiss it with a reason
`mark_implemented`	Yes	Record that the change was made, with a note; starts the 28-day review and annotates the day
`get_outcomes`		Implemented actions with their reviews (before, after, verdict), and ones still waiting
`list_markers`		Annotations: releases, notes, team-wide ones and Google updates
`create_marker`	Yes	Annotate a release or change; with a `ref`, calling again updates it
`delete_marker`	Yes	Delete an annotation on a site
`list_experiments`		Experiments with their cached before and after
`get_experiment`		One experiment, analysed now
`create_experiment`	Yes	Start measuring a change to a page, a group of pages or the whole site
`list_url_groups`		Saved URL groups with totals for a period against the one before
`create_url_group`	Yes	Save a URL group from rules (contains, starts with, exact, regex) and excludes
`list_alerts`		Recent alerts for a team, newest first

Search Console data runs about two days behind. list_properties tells the assistant the last complete day for each site, and gsc_metrics never reads past it, so a "last 28 days" answer is 28 days of finished data rather than 26 and two empty ones.

Query-level rows leave out the queries Google anonymises, so totals built from query text come out lower than page totals. The assistant is told to expect the difference.

How do you connect it?

Claude Code, Claude.ai and Claude Desktop

These sign in with OAuth. Sign-in only sends you back to Claude's own addresses or to a program on your computer (such as Claude Code or mcp-remote), so a look-alike app can't collect your approval. In Claude Code:

claude mcp add --transport http copilot-seo https://copilot-seo.com/mcp

Then run /mcp and sign in through the browser. In Claude.ai or Claude Desktop, add the same URL as a custom connector. You approve a consent screen that names the client, after confirming your password.

Access tokens last a day and the client refreshes them itself. Refresh tokens last 90 days. A client that asks for read access gets a read-only token.

Scripts and other MCP clients

Any other client connects with a personal access token instead. Create one on the Connect AI assistant page in Settings. Tokens can always read. They can only change anything if you tick "Allow writes" when you create them.

curl -fsS https://copilot-seo.com/mcp \
  -H "Authorization: Bearer $COPILOT_SEO_TOKEN" \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"list_properties","arguments":{}}}'

What can a token see?

Every tool acts as the user who owns the token, across the teams that user belongs to. A site, annotation or experiment outside those teams comes back as not found. Tools that change anything refuse a read-only token. The Connect AI assistant page lists connected apps as well as tokens, and disconnecting an app ends its access straight away. Requests are limited to 120 a minute per user.

Query text, page URLs and notes can come from anyone: searchers, other sites, your teammates. The server tells the assistant to treat them as data, never as instructions.

Where it fits

The MCP server is a way into the rest of Copilot SEO, not a separate product. The Action plan it reads ranks CTR opportunities, keyword cannibalisation and content decay alongside each other, and create_experiment starts the same SEO experiments you can create in the app. A useful pattern is to have your deploy pipeline or your assistant add an annotation for every release, so a change in the figures can be read against what shipped.

There is no tool to end or cancel an experiment yet. Do that in the app.

An assistant can start from the guide for agents. Once it is connected, it also gets the full guide, with each routine step by step, and a ready-made prompt for each routine.

A Google Search Console MCP server