TOOLKIT
Google Analytics MCP
Google’s own MCP server for GA4: seven read-only tools over the Admin and Data APIs, so an agent can run reports on your property without a CSV export.
by Google (official)googleanalytics/google-analytics-mcpApache-2.0 licence3.2k starscommit a8ca729
Marked experimental by Google and published on PyPI as `analytics-mcp`. Seven tools: account summaries, property details, Google Ads links, a core report runner, a funnel report runner, custom dimensions and metrics, and a realtime report.
Read-only by construction: the credentials must carry the `analytics.readonly` scope and nothing else. There is no tool that writes to a property.
The setup is the cost: a Google Cloud project with the Admin and Data APIs enabled, Application Default Credentials for a user with access to the property (or a service account you impersonate), and pipx to run the server.
What it exposes
7 tools, and which ones change things
All tools: get_account_summaries get_property_details list_google_ads_links run_report run_funnel_report get_custom_dimensions_and_metrics run_realtime_report
Writes
None. There is no tool in this server that changes your data.
Pairs with
Skills that work on what it reads
- Tracking plan and GA4 setup
analyticsEvents, GA4, Tag Manager and UTMs, planned before any number is read. - Attribution models
attributionWhich channel caused the sale: seven models, how each lies, and how to reconcile. - Reading the ads dashboard
ads-performance-analyticsPlatform-reported versus real, reconciliation, ROAS versus LTV, incrementality.
How to install these skills: three routes into Claude Code, and one line to check they loaded.
Method behind it
Once the server runs, the questions worth asking it are the ones our ROAS glossary sets up: revenue per channel over a fixed window, compared with what the ad platform claims for the same window.
What is ROASQuestions
Questions people ask before installing
Can it change anything in my GA4 property?
No. The scope is read-only and the tool list has no write. The worst a bad prompt can do is run an expensive report.
Does it work with Claude Code as well as Gemini?
Yes. The README documents both: a settings.json entry for Gemini and a `claude mcp add` line for Claude Code, with the same two environment variables.
Whose data does it see?
Whatever the credentialed user or service account can see in GA4. Use a user with access to only the properties you want the agent to read.