GAGA4 MCPv2.11.9

v2.11.6 · MIT · local stdio · read-only

Ask GA4 a question. Get the answer an analyst would give.

One MCP server for Claude, Cursor, Codex, Gemini CLI and fourteen more clients. Twelve tools, sixteen analytical skills, and a schema check that validates every field name against your property before the query runs.

Install

shell
curl -fsSL "https://ga4.builditwithai.xyz/install" | bash

Finds your client, writes the entry, checks the credentials.

12
tools
16
analytical skills
17
client guides
0
writes to your property

An agent with a GA4 tool guesses. That is the whole problem.

Three failures show up every time someone points a general-purpose model at the Analytics Data API.

It invents field names

Models are trained on Universal Analytics vocabulary. pageviews instead of screenPageViews. The API returns a 400 and the agent retries with another guess.

It invents a method

Ask why traffic dropped and it will produce a plausible five-step analysis that never checks whether the drop was one channel or all of them.

It does the arithmetic itself

Pull seven daily rows, ask for the week, and the model sums them in its head. Rates do not sum. The number is wrong and looks fine.

What this server does about it

Three mechanisms, all of them inside the tool call. Nothing here depends on the model behaving well.

Schema check before the query

get_ga4_data validates every dimension and metric name against your live property. A wrong name comes back with the reason and the correct spelling, not a 400 and a retry loop.

Skills loaded on demand

Sixteen recipes ship in the package — traffic diagnosis, ecommerce, attribution scope, AI referral split. The agent loads one before it analyzes, so the method is fixed instead of improvised.

Totals computed by GA4

Every multi-row pull returns a totals block that GA4 computed server-side. The model reads the period figure instead of adding up rows and drifting.

What it does when the setup is broken

Most failures are not analysis failures. They are a missing property ID, an expired token, or a service account that was never added to the property. The server handles all three inside the session.

Setup recovery mid-session

setup_ga4_access collects the missing value and reconnects without a restart. No client config editing, no reinstall.

Guides bundled offline

get_troubleshooting_guide serves the setup, IAM and schema fix paths from inside the package. It works with no network and no API access.

Row guard before the pull

Large queries are estimated first and warn above 2,500 rows, so one careless question does not fill the context window with raw rows.

Against Google's own server

Google ships an official Analytics MCP server, and it is genuinely good. Here is the honest split.

Google's Analytics MCP

  • Admin API plus Data API: accounts, properties, Google Ads links, core and funnel reports, realtime.
  • Hosted remote endpoint at analyticsdata.googleapis.com/mcp/v1 — no local install at all.
  • Apache-2.0, labelled Experimental, maintained by the Analytics organisation.
  • Best choice when you want zero install and a vendor-maintained toolset.

This server

  • Reporting and metadata, with the analysis layer attached: 16 skills, schema validation, GA4-computed totals, row guard.
  • Local stdio over your own credentials. Nothing leaves your machine except GA4 API calls.
  • MIT, one maintainer, and a fix path for the three errors that actually stop people.
  • Best choice when the agent has to reason about the numbers, not just fetch them.

Natural language analytics & production prompts

Query your analytics data directly from conversation without memorising API dimension names or building filter JSON by hand.

How natural language GA4 works

Why raw models hallucinate Universal Analytics metrics, how the MCP schema layer bridges the vocabulary gap, and why local stdio keeps company traffic numbers private.

35 copy-paste analytical prompts

Production prompts for traffic diagnosis, ecommerce funnels, AI referral tracking, content audits, and bot traffic detection across Claude, Cursor, and Gemini.

Sixteen skills, loaded only when needed

Each one is a short playbook with the exact dimensions, metrics and interpretation rules for one job. The agent calls search_skills, gets one back, and follows it.

SkillWhat it settles
Traffic diagnosisWhether a change is real, and which channel moved first
EcommerceRevenue, conversion rate, AOV and funnel drop-off
AI referralsSplitting ChatGPT, Claude, Perplexity and Gemini traffic from organic
Attribution scopeWhy session, event and user numbers never match
Metric namesThe GA4 name for the UA name the model guessed
All sixteenIncluding filters, date ranges, custom dimensions, bot traffic

Your client is on the list

The installer writes the entry for eight of them. Codex, Gemini CLI, Goose and Kiro take one config block each, quoted from that vendor's own documentation.

Auto-configured

Claude Desktop and Code, Cursor, VS Code (Cline, Roo Code), Continue.dev, OpenCode, Devin Desktop (formerly Windsurf), Zed, Google Antigravity.

One block, copied from the vendor

Codex writes TOML, Goose writes YAML, Gemini CLI and Kiro write JSON. Each guide shows the exact shape that client expects.

Start here

The client hub has every guide, or go straight to Claude or Cursor.

Install

  1. Run the installer

    shell
    curl -fsSL "https://ga4.builditwithai.xyz/install" | bash

    It detects your client, checks for uv, and writes the server entry. Restart the client afterwards.

  2. Give it two values

    shell
    export GA4_PROPERTY_ID="123456789"
    export GOOGLE_APPLICATION_CREDENTIALS="/absolute/path/to/key.json"

    Both live inside the client's server entry. The property ID is the numeric one from Admin, Property details — not the G- measurement ID.

  3. Ask a question that needs a real answer

    text
    Why did organic sessions drop in the last 7 days?

    You have succeeded when the agent loads the traffic-diagnosis skill and calls get_ga4_data without a setup error. Full setup guide.

The parts people get stuck on

Error triage & fixes

Step-by-step triage for 403 permission denied, 400 invalid field names, 429 quota limits, and scope errors.

The 403 that is not a bug

The service account is valid but was never added to the property. Nine steps in the Analytics UI, with the exact role.

The filter shape

dimension_filter follows the Data API v1beta schema, not SQL. Working templates for string, list, numeric and AND/OR groups.