Works with the AI Agents you already use
GA4 MCP Skills for AI Coding Agents
Stops your agent guessing GA4 field names. Answers analytics questions with the right tool, then builds the tracking plan your reports depend on.
$4.20
GA4 MCP Skills for AI Coding Agents
Example session with this skill installed
Sessions and key events by channel for the last 28 days vs the previous 28, for Acme (property 4815162342). Flag anything that looks like a tracking problem, not a demand problem.
- Read your context and instructions
- Compiled the ga4 mcp
Step 1 — resolve the field name before spending a call.
get_metadata(property_id="4815162342", search="key event")
{"metrics": [
{"api_name": "keyEvents", "ui_name": "Key events", "category": "Conversions",
"type": "TYPE_INTEGER"}
]}
keyEvents confirmed. conversions is not on this property. No failed call needed to find that out.
Step 2 — check the scope combination before running it.
The agent wanted sessionDefaultChannelGroup + eventCount. Session-scoped dimension, event-scoped metric: one session with several events counts several times. It reports sessions and keyEvents instead, which are both session-context metrics.
Step 3 — one report, two named date ranges.
run_report(
property_id="4815162342",
dimensions=["sessionDefaultChannelGroup"],
metrics=["sessions", "keyEvents"],
date_ranges=[
{"start_date": "28daysAgo", "end_date": "yesterday", "name": "current"},
{"start_date": "56daysAgo", "end_date": "29daysAgo", "name": "previous"}
],
order_bys=[{"metric": {"metric_name": "sessions"}, "desc": true}],
metric_aggregations=["TOTAL"]
)
The result
Sessions and key events by default channel group Property 4815162342 · 2026-10-19 → 2026-11-15 vs 2026-09-21 → 2026-10-18 · time zone Europe/London
Channel Sessions Prev. Δ Key events Prev. Δ Key event rate
Organic Search 18,420 16,905 +9.0% 1,284 1,196 +7.4% 6.97%
Direct 9,880 10,412 −5.1% 642 705 −8.9% 6.50%
Paid Search 6,340 7,105 −10.8% 388 562 −31.0% 6.12% ↓
Organic Social 2,150 2,890 −25.6% 61 88 −30.7% 2.84%
Referral 1,205 1,140 +5.7% 47 41 +14.6% 3.90%
Email 980 610 +60.7% 214 96 +122.9% 21.84%
Unassigned 42 38 +10.5% 2 1 — —
The read-out
Organic Search is healthy: sessions +9.0%, key events tracking with them. Email is a campaign launch, not a trend — 610 → 980 sessions and key events more than doubled.
Paid Search is not a demand problem. Sessions fell 10.8%, but key events fell 31.0%, and the key-event rate dropped from 7.91% to 6.12%. Traffic does not fall three times faster than the conversion rate on its own. Something stopped counting on paid landing pages, in the last two weeks.
Two candidates worth checking: a tag change, or a consent change on the paid landing template. I can confirm which one if you give me the GTM container export — that's a five-minute audit.
Connects securely to your tools. The creator never sees your data.
About this skill
What this is
Two agent skills that make Google Analytics 4 usable by an AI coding agent.
Your agent can call the GA4 API. What it can't do is know that the metric is
keyEvents and not conversions, that screenPageViews is not pageviews,
that returnOnAdSpend is not returnOnAdSales, that session-scoped metrics
don't sum across page dimensions, or that funnels simply aren't available
through this API. So it guesses. The call fails, or worse, it succeeds and
returns a confident wrong number.
These skills encode the difference between an agent that has GA4 access and an agent that can answer a GA4 question correctly.
The two skills
ga4-mcp — the query path
Fires when a request needs real GA4 numbers or configuration details: traffic, users, sessions, engagement, key events, revenue, ecommerce, channels, campaigns, landing pages, geography, device, new vs returning, realtime activity, custom events, streams and measurement IDs.
It contributes:
- A routing table from question to tool, across all ten tools the server exposes.
- The call order that avoids wasted calls: exact field names first, compatibility check second, one report third.
- The full reporting contract — date tokens, filter shapes, ordering, comparisons, aggregation, pagination, quota response.
- Method rules that keep numbers honest: which metric people actually mean, where totals silently double-count, when a window is partial, which fields must never be mixed.
- Twelve copy-ready report payloads and a field cheatsheet.
ga4-tracking-plan — the documentation path
Fires when there is no tracking plan, or one needs creating or auditing.
It reconciles two versions of the truth: your Google Tag Manager container (the events that should be firing) and live GA4 data (the events that are). Every event lands in exactly one evidence class — tag and data, tag only, data only, or proposed — never in a bucket chosen by vibes.
It ships a deterministic parser for GTM container exports. Do not point an agent at a 20,000-line export and hope. The parser extracts GA4 event tags, their parameters and user properties, the triggers that fire them, the Data Layer variables behind them, and the defects already in the container:
- the same event name sent by two different tags
- tags that never fire because no trigger is attached
- events wired to All Pages, inflating on every route change
- event names that are unresolved
{{variables}} - parameters that look like personal data
- event and parameter names past Google's 40-character limit, which GA4 silently drops
Output is docs/analytics/tracking-plan.md: event catalog, parameter registry
with scopes and cardinality, key events with counting rules, gaps, PII flags,
and change log. That file then becomes the input the query skill reads, so
"signups" maps to sign_up with method and plan_name — not a guess.
Why the plan matters more than the query
Every analytics question that goes wrong in practice goes wrong for the same reason: two people meant different fields. The plan fixes that once. It also catches the failure mode nobody notices — an event that was implemented, reviewed, and shipped, and has never fired a single time.
Requirements — read this before buying
This is a skills package, not a hosted service. It needs:
- A
ga4MCP server connected to your agent. Free and open source (github.com/furiber/ga4-mcp). Run it locally, or host it on Render, Google Cloud Run or Cloudflare. Sign in with your own Google account; it requests onlyanalytics.readonly. - Viewer access to the GA4 property you want to report on.
- Python 3.9+ in the agent's shell — standard library only, no packages, for the GTM export parser.
- Optional but strongly recommended: a GTM container export (
.json) for the tracking plan. Without it, the plan still builds from observed data, but it cannot state triggers, and it says so in the output.
Buy this if you already run the ga4-mcp server, or you are willing to set it up in about ten minutes. Don't buy this if you want a hosted GA4 connector with no setup — this won't give you one, and the skills are honest about it.
What it does
- Picks the correct MCP tool for a question and says which one it used
- Produces valid report payloads, including filters, comparisons and paging
- Separates metrics that look interchangeable and aren't
- Flags partial windows, aggregated rows, truncation and quota pressure
- Builds a tracking plan with an explicit evidence trail per event
- Audits instrumentation: declared-but-silent, arriving-undocumented, duplicated, misconfigured
- Writes the
dataLayer.pushcontract for proposed events - Recovers from the common server errors with the specific fix
- Stays inside read-only scope at all times
What it does not do
Stated plainly, because a refund argument starts with "the page didn't say".
| Not supported | Why |
|---|---|
| Writing to GA4 | Read-only scope, enforced by the API |
| Funnel reports | No funnel endpoint exists in this API |
| Cohort and retention curves | No cohort report support |
| Reading your retention setting | Not exposed by the property API; it asks instead |
| Search Console, Ads platform, BigQuery, CRM | Different data sources |
| Creating or publishing GTM tags | Produces a specification, never touches your container |
| Proving a tag fired on a specific page | Needs a debug session or BQ export |
| Dashboards and scheduled reports | Output is Markdown and chat tables |
| Fixing broken instrumentation | It diagnoses; a human ships the change |
| Auditing several GTM versions at once | One export per run, version recorded in the plan |
Notes on accuracy
Field names, filter enums, character limits and custom-dimension budgets were verified against the Google Analytics Data API, Admin API and Tag Manager API discovery documents, not from memory. Every claim of "not supported" above was checked the same way. Where something varies — GTM export dialects, undocumented tag codes, per-property field drift — the skill says so and tells the agent to verify against the live property rather than assume.
The v1.0.0 changelog ships with the package and lists the corrections made during review, including a dozen facts that were wrong in draft.
What to ask it
- "Sessions and key events by channel, last 28 days vs the previous 28."
- "Top landing pages by engaged sessions last month, France only."
- "Is anyone on the site right now, by country?"
- "Why did our conversion count drop last week?"
- "Create a tracking plan from our GTM export and the live property."
- "Is
account_createdactually firing, or is it only in the container?"
Package contents
- 2 skills, 9 files
- Entry documents kept under 165 lines each; detail loads on demand so context stays cheap
- A GTM export parser (standard library only)
- A changelog, an installer for macOS/Linux and Windows, and a README that documents limits as carefully as features
Works with Claude Code, Cursor, Codex CLI, Gemini CLI, GitHub Copilot, Windsurf, Cline and any agent that loads SKILL.md. One purchase, all of them.
v1.0.0 · free updates as GA4's field names change.
Read-only access only. Cannot edit GA4 config or GTM tags. Requires GTM JSON export for full auditing. No support for Search Console or Ads data directly.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 1 day ago
Requires