The short answer
Pick Google's Analytics MCP server for deep GA4 work on a handful of properties: funnels, realtime, exact dates and property settings, if you can run Python, gcloud and a Cloud project. Pick a hosted, stored connector when you compare many sites, need Search Console next to GA4, or want ChatGPT without hosting anything. You can also run both.
- Google's server is GA4 only, live, and runs on your machine or a server you operate
- A hosted connector trades GA4 depth for breadth: more sources, more sites, nothing to install
- Running both side by side works if you tell your agent which questions go where
Side by side: what each one does
Start with the shape of each tool. They solve different problems, and the rows where one clearly wins decide most setups.
| Question | DuckDive (hosted) | Google's server |
|---|---|---|
| Where it runs | On DuckDive's servers; your client gets a single address | Your computer, or a server you run |
| What you install | Nothing, sign in with Google | Several things: Python, pipx, gcloud, Cloud project |
| Data sources | Six: GA4 and Search Console, Google Ads and LSA, Yext, Local Falcon | GA4 only, plus its Google Ads links |
| GA4 report depth | Partly: totals, channels, landing pages | Yes: standard, funnel, conversions, realtime |
| Date ranges | Fixed windows, 7 days to 12 months | Dates you name |
| Many properties in one question | Yes, each site already imported | Partly, a live report per property |
| Changes to GA4 settings | No | Opt-in: key events, dimensions, annotations |
| ChatGPT | Yes, on Business or Enterprise | Only after you host it at a URL |
| Price | Free, open source (Apache-2.0); you pay for any hosting |
Sources: Google's Analytics MCP server guide, the google-analytics-mcp README and release notes, checked November 2026. For the DuckDive column, see how DuckDive connects to AI apps.
What Google's server is good at
As published in the project README (checked November 2026), the server sorts its tools into three groups. Most of this is GA4 depth that a breadth-first connector leaves out.
- Account and property information.
get_account_summaries,get_property_details,list_google_ads_links,list_property_annotationsandlist_key_events. Your agent can audit how a property is configured, not just what it reports. - Core reports.
run_reportfor standard reports,run_funnel_report,run_conversions_report(the README describes it as covering ad costs, conversions, ROAS and attribution) andget_custom_dimensions_and_metrics. - Realtime.
run_realtime_report. Version 1.1.0 added minute ranges to it, according to the release notes.
Funnels are the clearest win
If your question is about steps inside one visit, Google's server is the right tool and DuckDive is not. DuckDive has no funnel report. Its GA4 tools stop at daily totals, channel splits and landing pages. Note one detail from the release notes: the funnel tool arrived in 0.3.0 and is built on the Data API's v1alpha method, so expect its behaviour to shift as Google changes that API.
Ask your agentShow me a funnel from product page to checkout for last week.
Tools it calls run_funnel_report on Google's server. This one is not a DuckDive job.
Auditing how a property is built
Before anyone trusts a quarter of numbers, it helps to know what the property actually counts. The account and property tools let an agent read that configuration directly: which events are marked as key events, which custom dimensions and metrics exist, and which Google Ads accounts are linked. DuckDive has nothing like this. It reads reported figures, never the setup behind them.
Ask your agentAudit the Wickerbay Rugs GA4 property before Thursday's review: list every key event with its counting method, every custom dimension and metric, and the linked Google Ads accounts. Flag any key event that has not fired once in the past four weeks.
Tools it calls get_property_details list_key_events get_custom_dimensions_and_metrics list_google_ads_links run_report on Google's server. None of these exist in DuckDive.
A stale key event, say a "newsletter_signup" left over from an old theme, quietly inflates every conversion rate built on it. Catching that is configuration work, and it belongs to Google's server. Because the same server queries GA4 live, it also accepts any calendar dates you name (a promo weekend, the fortnight before a relaunch) and can read the last few minutes of traffic with its realtime tool. DuckDive's agent tools work from imported history in preset windows.
Read-only, with a write switch
Google's developer guide says the server "is available for read requests only." The README now goes further than the guide. Since 0.8.0 there is an --allow-write flag, off by default. Turn it on, add the analytics.edit scope to your credentials, and the agent can create annotations, and since 1.1.0 also key events and custom dimensions. So "read-only" is a default: check which flags your config passes.
What setting Google's server up involves
No single step is hard for a developer, but together they assume a working Google Cloud setup. The trade press said the same about 1.0: The PPC Times wrote that "Setup is still not one click." The path below follows Google's guide, checked November 2026.
-
Install Python 3.10 or newer, plus pipx
The package is
analytics-mcpon PyPI. Most clients start it withpipx run analytics-mcp. Google's Cursor example usesuvx --from analytics-mcp google-analytics-mcpinstead. -
Pick a Google Cloud project and enable two APIs
Turn on the Google Analytics Admin API and the Google Analytics Data API in that project. Your client config later names the project in
GOOGLE_CLOUD_PROJECT. -
Create Application Default Credentials
Sign in through gcloud with the read-only Analytics scope and the cloud-platform scope. Service account impersonation is the documented alternative.
$ terminal gcloud auth application-default login --scopes="https://www.googleapis.com/auth/analytics.readonly,https://www.googleapis.com/auth/cloud-platform"
-
Check Viewer access on every property
The signed-in identity needs at least the Viewer role on each GA4 property you plan to ask about. If you use a service account instead, its email address has to be added to each property as a user.
-
Register it in your client
Google's Claude Desktop example runs
pipxwithrun analytics-mcpand passesGOOGLE_APPLICATION_CREDENTIALSandGOOGLE_CLOUD_PROJECTas environment variables. Cursor, Claude Code, Claude Desktop and Antigravity each get their own section in the guide, and the Gemini CLI section is marked deprecated.
Self-hosting for a team
Version 1.0.0 added a multi-tenant streamable HTTP mode with OAuth 2.1, so a team can share one deployment while each person signs in as themselves. Version 1.1.0, three days later, added a Dockerfile and Cloud Run deployment docs, per the release notes. A deployment like that is also what ChatGPT needs: OpenAI's developer mode guide connects remote servers only, so a copy running on a laptop is invisible to it.
Before you deploy
The README says the built-in OAuth provider keeps client registrations, pending sign-ins and refresh tokens in process memory. Run it as a single instance or behind sticky sessions, and expect everyone to sign in again after a restart.
Where a hosted, stored connector differs
DuckDive copies the history from each linked source and holds on to it. On the agent side there is a single endpoint, https://api.duckdive.ai/mcp, reached through OAuth. Nothing gets installed, no Cloud project is involved, no key file sits on a laptop, and the agent is handed a DuckDive token rather than any Google password.
The stored part is what changes the answers. A question about nine sites becomes nine reads of data that is already there, instead of nine live reports against Google. Google counts Data API usage per property and per project, so live tools feel that cost first on long windows and many properties. We work through that arithmetic on imported data against live MCP calls rather than repeat it here.
What the hosted side gives up
Be clear about the cost of breadth before you choose it. DuckDive's GA4 tools return three things: daily sessions and conversions with totals and rate (get_ga4_overview), the channel or source split (get_ga4_channels), and top landing pages (get_ga4_landing_pages). Each call covers a single website.
- Nothing on funnels, live visitors, revenue, audiences, or splits by device, country or event parameter.
- No calendar dates of your choosing: the shortest preset is a week and the longest a year.
- Landing page, query and page lists stop at 100 rows a call.
- No tools for GA4 configuration. Marking an event as a key event stays in GA4 Admin, or with Google's server once writes are switched on.
The full GA4 metric list, with definitions, is on our GA4 integration page.
What the stored side is for: a nine-property ranking
The fair test is a question both tools can answer, like ranking conversions across nine properties.
Ask your agentCompare conversions across my nine properties over the last 6 months and rank them by change.
Tools it calls list_websites get_ga4_overview
Compare conversions across my nine properties over the last 6 months and rank them by change.
list_websitespulls the nine Wickerbay Home sites and checks each has a GA4 property attachedget_ga4_overviewnine calls, one per site, asking for a year of daily rows so the recent half can be set against the earlier half
I took "over the last 6 months" to mean the newest half-year set against the half-year before it. If you wanted the slope within the newest half-year, say so and I will redo it.
| Site | Latest 6 months | Change |
|---|---|---|
| Wickerbay Outdoor | 1,012 | +58.1% |
| Wickerbay Outlet | 803 | +14.4% |
| Wickerbay Kids | 471 | +9.3% |
| Wickerbay Bedding | 1,515 | +8.1% |
| Wickerbay Trade | 255 | +3.2% |
| Wickerbay Rugs | 2,258 | +2.0% |
| Wickerbay Kitchen | 987 | +0.6% |
| Wickerbay Journal | 112 | -5.1% |
| Wickerbay Lighting | 1,703 | -7.9% |
Outdoor tops the list, but that is the calendar: the latest 6 months hold the summer season and the prior 6 hold winter. Set it aside and Outlet leads. Lighting deserves the attention: it ranks second for conversions and is the only large site shrinking.
Google's server answers this too, with exact calendar dates as a bonus: get_account_summaries finds the nine property IDs, then run_report runs live for each. For an occasional question that works well. It gets heavier as properties, windows and weekly repeats grow.
Worked example: one question asked both ways
Now a question that crosses sources, the kind an operator with a family of sites asks after a soft quarter.
Ask your agentFor Wickerbay Home's nine sites, which one lost the most Search Console clicks in the last 90 days against the 90 before, and did its GA4 conversions move with it?
Tools it calls list_websites get_search_console_summary get_ga4_overview get_ga4_channels
Through DuckDive
Across all nine Wickerbay properties, which property shed the most organic clicks over the recent 90-day stretch versus the stretch before it, and did conversions drop alongside?
list_websitesreturns the nine sites and shows both Google sources attached to every oneget_search_console_summarya six-month daily series per site, cut at day 90 into a recent block and an earlier blockget_ga4_overviewthe same cut, applied to daily conversionsget_ga4_channelsonly for Wickerbay Lighting, pulling both preset windows to isolate organic conversions
Wickerbay Lighting lost the most: clicks fell from 39,600 to 27,300, down 12,300. Its GA4 conversions barely moved, 862 to 841. Changes below compare the two 90-day periods.
| Site | Clicks change | Conversions change |
|---|---|---|
| Wickerbay Lighting | -31.1% | -2.4% |
| Wickerbay Outdoor | -29.0% | -34.1% |
| Wickerbay Journal | -9.8% | -3.5% |
| Wickerbay Kitchen | -3.7% | -1.8% |
| Wickerbay Trade | -2.8% | +2.4% |
| Wickerbay Rugs | -2.7% | -1.9% |
| Wickerbay Bedding | +2.2% | +3.6% |
| Wickerbay Kids | +2.6% | +3.9% |
| Wickerbay Outlet | +6.3% | +7.0% |
Lighting's lost clicks were mostly not buyers. Its organic search conversions held at about 390 in both periods. That points at informational queries slipping, not product pages. Pull its top Search Console pages next to see which ones.
Outdoor is the opposite pattern: clicks and conversions fell together, which fits the end of the outdoor season. Check the same months last year in Search Console itself, because my windows stop at 12 months.
One caveat on Lighting: its earlier organic number comes from subtracting the 90-day figure from the half-year figure, and the half-year spans slightly more than 180 days, so read it as a close estimate.
Through Google's server
The same prompt, sent to an agent that only has Google's Analytics server connected, runs like this:
get_account_summarieslists the accounts and properties, and the agent matches nine of them to the Wickerbay sites by name.- The Search Console half has nowhere to go. The server has no Search Console tools, so a careful agent says so. A careless one guesses from GA4's organic channel, which is not the same number.
run_reportruns for each of the nine properties, covering both 90-day periods with exact dates, each one a live Data API request.- The agent can rank conversion changes, but it cannot say which site lost clicks, which was the actual question.
To finish the job you would add a separate Search Console server alongside it; our Search Console integration page covers the options. Google's half would still be sharper on one thing: if you then wanted to know whether Lighting's product pages lost add-to-carts, run_funnel_report can show it and DuckDive cannot.
Get the clicks and conversions answer for your own sites
Attach Search Console and GA4 to every site one time, and the cross-site clicks question becomes a single prompt.
Free for 2 sites · No credit card
Running both at once
You do not have to choose forever. A single MCP client will load more than one server, and these two never collide on a tool name. Google's include run_report and get_account_summaries; DuckDive's include get_ga4_overview and get_search_console_summary. Keep Google's server for depth on your main property and let DuckDive carry the breadth. Here is the setup in Claude Code, where both are documented.
-
Leave Google's server where it is
If it already answers in your client, change nothing. Confirm with
claude mcp listthat it shows up. -
Add DuckDive with one command
This is the command from DuckDive's agent guide. No header, no API key.
$ terminal claude mcp add --transport http duckdive https://api.duckdive.ai/mcp
-
Sign in from inside Claude Code
Type
/mcp, pick DuckDive, and select Continue with Google on the page that opens. If you want the agent to create sites on your behalf, keep the option that permits actions switched on; only admins see it. -
Have the agent link your sites
The agent reads the sites your Google login can see, creates them in DuckDive, attaches whatever sources line up, and checks with you before reassigning a property already in use. For the long version, see agent-led DuckDive setup.
Paste into Claude CodeSet up DuckDive for my website. Follow https://analytics.duckdive.ai/
agents.md -
Write a routing note
Put a short rule in your project instructions so the agent stops guessing which server to use. Adjust the wording to your own sites.
Ask your agentAdd this to our project notes: route multi-site questions, and anything touching Search Console, paid search or LSA leads, to the DuckDive tools. For funnels, realtime visitors, exact calendar dates or property settings on a single GA4 property, use the Google Analytics server. If a question needs both, say which part came from which.
-
Confirm the history landed before you rely on it
An answer from stored data can only be as current as the last import. Once linking is done, have the agent report on each site.
Ask your agentFor every Wickerbay site in DuckDive, confirm that analytics and search data are each hooked up, whether each import finished, is running or errored, and the newest date each one holds.
Tools it calls
list_websitesget_sync_status
Claude Code specifics, including permission rules for the DuckDive tools, are on our Claude Code setup page. In ChatGPT the picture is different: DuckDive connects as a custom MCP server on Business or Enterprise plans, while Google's server would first need a public HTTPS deployment. The steps are on our ChatGPT connection page.
Numbers may differ slightly
Google's server asks GA4 live; DuckDive reads its last import. Expect small gaps on the latest day or two, and check get_sync_status for how fresh DuckDive's copy is.
Which to choose, by situation
Match your most common question to a row. Split across both columns? Run both.
| Your situation | DuckDive | Google's server |
|---|---|---|
| Drop-off between product page and checkout | No, no funnel tool | Better fit, run_funnel_report |
| Cleaning up key events before a review | No, cannot read configuration | Better fit, property and key event tools |
| A promo weekend with exact dates | Partly, preset windows only | Better fit, any dates |
| Adding key events, dimensions or annotations | No, no GA4 settings tools | Better fit, with writes turned on |
| A ChatGPT workspace and no server to host | Better fit, public URL on Business or Enterprise | Partly, after you deploy HTTP mode |
| Nine or ninety properties, every week | Better fit, history already on hand | Partly, a live report per property |
| Search Console beside GA4 | Better fit, both linked per site | No, GA4 only |
| Ads, LSA or local rankings too | Better fit, six sources | No, outside its scope |
| Nobody wants to run Python or gcloud | Better fit, one URL and Google sign-in | Partly, someone has to maintain it |
| No third-party copy of your data | No, storing a copy is the point | Better fit, stores nothing itself |
Sources: google-analytics-mcp README and Google's Analytics MCP server guide, checked November 2026.
Five questions to settle it in ten minutes
- Count your properties. If you ask about one or two, Google's server alone is often enough.
- List your last ten questions. Mark each one as depth inside a visit or breadth across sites and sources.
- Name your client. ChatGPT users need a URL; Claude Code and Cursor users can run either.
- Find your gcloud owner. If nobody on the team will keep credentials and a Cloud project healthy, hosted wins.
- Ask where data may live. If policy forbids a third-party copy, stay with Google's server and accept live queries.
No MCP server at all? The cross-property GA4 guide for standard accounts walks through exports and roll-up alternatives. For community and hosted servers beyond these two, see marketing data MCP servers compared.
Questions about the official server
What changed when Google's Analytics MCP server reached 1.0?
Version 1.0.0, released November 3, 2026, moved the project out of experimental status and added a multi-tenant streamable HTTP transport with OAuth 2.1. Version 1.1.0 followed on November 6 with opt-in write tools for custom dimensions and key events, minute ranges for realtime reports, and Docker and Cloud Run deployment docs. PyPI now labels the package Production/Stable.
Is Google's Analytics MCP server read-only?
By default, yes. Google's guide says it serves read requests only. The README adds an --allow-write flag, off unless you pass it, that enables tools to create annotations, key events and custom dimensions. Those also need the analytics.edit scope on your credentials, so without that scope the write tools cannot succeed.
Does Google host the Analytics MCP server for me?
No. The guide has you install and launch it locally through pipx or uvx. Version 1.0.0 added an HTTP mode with OAuth 2.1 so you can host it yourself, and 1.1.0 added a Dockerfile and Cloud Run instructions. Whoever deploys it owns the uptime, the upgrades and the sign-in state. With a hosted connector like DuckDive, that operator is someone else.
Can I use the official GA4 MCP server in ChatGPT?
Not as installed. ChatGPT only talks to an MCP server it can reach on the public internet, and Google's guide has no ChatGPT instructions, so you would first deploy the HTTP mode behind an HTTPS address. DuckDive is already served from a public URL, and ChatGPT accepts it on Business and Enterprise workspaces (Free and Plus cannot add it).
Does the official server read Search Console or Google Ads data?
No Search Console tools at all. On the Ads side it can list a property's Google Ads links, and its conversions report includes ad cost and ROAS as GA4 sees them. It does not read your Google Ads account. Next to GA4, DuckDive covers Search Console queries, a summary of each Google Ads account, LSA leads, Yext listing status and review counts, and Local Falcon map rankings.
Will running both servers confuse my agent?
Seldom. Their tool names never overlap and each tool carries its own description. Trouble starts on questions either side could answer, such as one site's conversions for October. A routing note in your project instructions settles those, and if the agent labels every figure with the server it came from, a mismatch stands out at once.
Which option keeps my analytics data in fewer places?
Google's server, run locally, keeps no copy: each report travels from Google to your client and on to whichever AI model you use. DuckDive holds an imported copy of every source you link for as long as the account stays open. In both setups the model's vendor decides what happens to tool output under its own terms, and DuckDive never sees the chat itself.
Sources
- Try the Google Analytics MCP server, Google for Developers, checked November 16, 2026.
- googleanalytics/google-analytics-mcp README, GitHub, checked November 16, 2026.
- Releases: googleanalytics/google-analytics-mcp, GitHub, checked November 16, 2026.
- Connect your AI assistant, DuckDive, checked November 16, 2026.
- Google Analytics MCP server reaches version 1.0 and leaves experimental status, The PPC Times, November 6, 2026, checked November 16, 2026.
- Add custom MCP server, OpenAI Developers, checked November 16, 2026.
- Limits and quotas on API requests, Google Analytics Data API, checked November 16, 2026.
- analytics-mcp, PyPI, checked November 16, 2026.
Last reviewed . Numbers in EXAMPLE DATA blocks are fictional.
