I look after SEO for five SaaS products. Every Monday I opened Google Analytics and Search Console ten times, once per product per tool, just to answer one question: what changed this week?
The obvious fix was to connect both to Claude. That is where I hit the two frustrations you have probably hit too:
The paid route. Hosted connector services do the Google side for you, but they charge a monthly fee for data that Google gives away for free.
The GitHub route. You copy someone's MCP server from GitHub, follow their README, sign in, and it works. Then a few days later it stops, and you sign in again. And again the week after.
The second one is not the repo's fault. It is one setting in Google Cloud that most READMEs skip, and I cover it in Step 5. Once that setting is right, you sign in once and the login keeps working.
So I built one dashboard that shows all five products in one place. It pulls data straight from the Google Analytics 4 and Search Console APIs, has an overview tab plus a tab per product, and refreshes itself every Monday. It took one evening. Most of that evening went into Google Cloud settings, not the dashboard.
This post is the full process, including every place I got stuck and exactly how I got unstuck.
The short version
- Create a Google Cloud project and enable three APIs: Search Console API, Google Analytics Admin API and Google Analytics Data API.
- Set up the Google Auth Platform, create a Desktop app OAuth client, and publish the app so your login does not expire every 7 days.
- Install two free, open-source MCP servers on your computer: Google's official GA4 server and a Search Console server.
- Connect both to the Claude desktop app, then ask Claude to pull the data and build the dashboard.
- Add a weekly scheduled task so it refreshes without you.
No paid connectors, no Looker Studio, and nothing gets billed to your Google Cloud account.
The four ways I could have built this
I looked at four setups before picking one. Only one of them costs money every month.
I went with the last option. A paid connector does the same job, but I did not want a monthly bill for free Google data. Looker Studio shows you numbers. With Claude connected to the data, I can also ask "why did impressions jump on this product?" and get an answer in the same place. That was the whole point for me.
Wait, what is an MCP server?
MCP (Model Context Protocol) is a standard way to give an AI assistant tools. An MCP server is a small program that runs on your computer and exposes a few actions, such as "run a GA4 report" or "list my Search Console sites".
Google publishes an official Google Analytics MCP server.
It is read-only by default, so it cannot change anything in your Analytics account. For Search Console I
used mcp-gsc by Amin Foroutan,
built by an SEO for SEOs. Its install name is mcp-search-console, which confused me
for a minute.
Why not just call the Google APIs from a web page?
That was my first idea. It does not work well, for two reasons.
- Sign-in breaks. Google's OAuth sign-in needs Google's own script, and hosted dashboard pages usually block outside scripts.
- Keys leak. A dashboard is a shareable link. Any API key or token inside it is visible to anyone who opens it.
With MCP, your Google login stays on your own computer. The dashboard only ever receives the finished numbers.
How it works
Google APIs
GA4 Data API
GA4 Admin API
Search Console API
MCP servers
analytics-mcp
mcp-search-console
hold your login
Claude
pulls 28-day data
checks the totals
runs every Monday
Dashboard
Overview tab
a tab per product
weekly snapshot
Google's APIs send data only to the two MCP servers on your computer, which hold your login. Claude calls those servers, checks the numbers, and writes them into the dashboard as plain data files.
What you need before you start
- A Google account that can see your GA4 properties and Search Console sites. Check which one by opening analytics.google.com and clicking your profile picture.
- A Google Cloud account (free, no billing account needed).
- The Claude desktop app. I used a Mac; Windows works with different file paths.
- About an hour, and Terminal open.
The setup, step by step
Not very technical? Do half, and let Claude do the rest.
- Steps 1 to 5 · You
These all happen in Google Cloud in your browser. It is just clicking, and Claude cannot do it for you because it needs your Google account. Do those yourself and keep the
client_secret.jsonfile from Step 4. - Steps 6 to 9 · Claude
Open Claude on your Mac, in a mode that can run commands on your computer (Claude Code works). Paste in this post (or its link) and send this:
I've done Steps 1 to 5 of https://www.pramanmenaria.com/blog/ga4-search-console-claude-dashboard and my client_secret.json is in Downloads. Do Steps 6 to 9 for me.
Claude installs the servers, creates the folder, runs the login and writes the config. You still do three
small things: pick your Google account when the browser opens, quit and reopen Claude, and check the test
in Step 9. Never paste the contents of client_secret.json into the chat. Claude
only needs to know where the file is.
In Google Cloud Steps 1 to 5, in your browser at console.cloud.google.com.
Create a Google Cloud project
Go to console.cloud.google.com, click the project picker at the top, then New project.
Name it SEO Reporting. No billing account is needed.
One name, three places
I use the same name, SEO Reporting, for every name Google asks for in this setup. Google asks three times, and one name means you always know which thing you are looking at.
- Step 1The project
- Step 3The app
- Step 4The client
Enable three APIs
Go to APIs & Services → Library, search for each one and click Enable:
Google Search Console APIYour search clicks, impressions and queries
Google Analytics Admin APILists your accounts and properties
Google Analytics Data APIRuns the actual reports
Google Analytics APINot this one. It is the old Universal Analytics API.
Where I went wrong
I first enabled only "Google Analytics API". That is the old Universal Analytics API. GA4 needs the Admin and Data APIs.
Set up the Google Auth Platform
In the left menu, open Google Auth Platform → Overview → Get started. Google walks you through four screens:
- 1 App information App name
SEO Reportingand your support email - 2 Audience Internal if you are on Google Workspace, otherwise External
- 3 Contact information Your email
- 4 Finish Accept the policy and click Create
If you chose External, open Audience → Test users and add the Gmail that owns your GA4 and Search Console data.
Create a Desktop app client
Open Google Auth Platform → Clients → + Create client.
- Application type
- Desktop app
- Name
SEO Reporting- Then
- Click Create, then Download JSON
Rename the file to client_secret.json. Treat it like a password: never paste its
contents into a chat or commit it to GitHub.
The first place I got stuck
The Clients page was empty and I thought something was broken. It is empty because no client exists yet; Google wants the consent screen (Step 3) set up first.
Publish the app, or your login dies in 7 days
This is the step behind the "why do I keep signing in again?" problem. An External app in Testing mode gives tokens that expire after 7 days. Most GitHub READMEs stop at Testing mode, so every week the tools quietly stop working and you sign in again.
Your first four weeks
Open Audience → Publish app → Confirm. The status should change to In production.
If Publish asks for a website and a privacy policy
That happened to me. The fix: go to Data Access, delete the two scopes you added, Save, then publish again. The scope list on that page is only used for Google's verification review. Your login still asks for the permissions it needs.
After publishing you will see a yellow banner saying "Your app requires verification". Ignore it. Verification matters for public apps. An unverified app works for up to 100 users, which is plenty for you and your team.
On your Mac Steps 6 to 9, in Terminal. If you are handing over to Claude, this is where it takes over.
Type yours and every path below updates to match, so you can copy the code as it is. Not sure what yours
is? Type whoami in Terminal. I left mine, praman, as the example.
Install the two MCP servers
I used uv, a fast Python tool installer. It downloads the right Python version for
you, so you do not have to manage one yourself. In Terminal:
# Install uv
curl -LsSf https://astral.sh/uv/install.sh | sh
export PATH="$HOME/.local/bin:$PATH"
# Google's official GA4 server
uv tool install --python 3.12 git+https://github.com/googleanalytics/google-analytics-mcp.git
# Search Console server
uv tool install --python 3.12 mcp-search-console
This puts two programs in ~/.local/bin: analytics-mcp and
mcp-search-console.
Sign in once for both tools
Both servers need a saved Google login. I wrote one small script that signs in once with both permissions and saves the login in the format each server expects.
First, make a private folder for everything to do with this setup. I named mine after the project,
seo-reporting, and put it inside ~/.config. The
~ is your home folder, so the full path is:
/Users/praman/.config/seo-reporting When you are done with this step, it will hold four files:
~/.config/seo-reporting/
├── client_secret.json ← the file you downloaded in Step 4
├── login.py ← the script below
├── ga_credentials.json ← created by the script, for GA4
└── gsc/
└── token.json ← created by the script, for Search Console
Make the folder and move your client file into it. This assumes you renamed the file to
client_secret.json and it is still in Downloads:
mkdir -p ~/.config/seo-reporting/gsc
mv ~/Downloads/client_secret.json ~/.config/seo-reporting/ Then save this script inside the same folder as ~/.config/seo-reporting/login.py:
import json, pathlib
from google_auth_oauthlib.flow import InstalledAppFlow
cfg = pathlib.Path.home() / ".config/seo-reporting"
secret = cfg / "client_secret.json"
SCOPES = ["https://www.googleapis.com/auth/analytics.readonly",
"https://www.googleapis.com/auth/webmasters"]
flow = InstalledAppFlow.from_client_secrets_file(str(secret), SCOPES)
creds = flow.run_local_server(port=0, access_type="offline",
prompt="select_account consent")
info = json.loads(secret.read_text())["installed"]
# GA4 server reads a standard Google credentials file
(cfg / "ga_credentials.json").write_text(json.dumps({
"type": "authorized_user",
"client_id": info["client_id"],
"client_secret": info["client_secret"],
"refresh_token": creds.refresh_token}))
# Search Console server reads its own token file
(cfg / "gsc" / "token.json").write_text(creds.to_json())
print("Saved. You can close the browser tab.") Run it from that folder:
cd ~/.config/seo-reporting
uv run --with google-auth-oauthlib python login.py A browser opens. Pick the Google account that owns your GA4 and Search Console data. You will see "Google hasn't verified this app"; click Advanced → Go to SEO Reporting. Tick both permission boxes and continue.
Two details that saved me later
select_accountin thepromptline forces the account picker, so you cannot sign in with the wrong Gmail by accident.- I keep the files in
~/.config, not in Documents, because macOS can block background apps from reading Documents.
Folders starting with a dot are hidden in Finder. To see .config, open your home folder and press Cmd Shift .
Wait, why does Search Console get full access?
I asked myself this too. The GA4 permission (analytics.readonly) is read-only. The
Search Console one (webmasters) is not, because the mcp-gsc server asks for full
access and will not run on a read-only login.
analytics.readonlyGA4: read onlywebmastersSearch Console: full, but risky tools stay off
In practice that is fine. The server's risky tools (adding or deleting a site, deleting a sitemap) stay
switched off unless you set GSC_ALLOW_DESTRUCTIVE to true. I
set it to false in the config below anyway, so nobody turns it on by accident.
Connect both servers to Claude
Open ~/Library/Application Support/Claude/claude_desktop_config.json (create it if
missing) and add both servers. The paths below use the username from the box above, currently
praman.
Change it if that is not yours.
{
"mcpServers": {
"google-analytics": {
"command": "/Users/praman/.local/bin/analytics-mcp",
"env": {
"GOOGLE_APPLICATION_CREDENTIALS": "/Users/praman/.config/seo-reporting/ga_credentials.json"
}
},
"search-console": {
"command": "/Users/praman/.local/bin/mcp-search-console",
"env": {
"GSC_OAUTH_CLIENT_SECRETS_FILE": "/Users/praman/.config/seo-reporting/client_secret.json",
"GSC_CONFIG_DIR": "/Users/praman/.config/seo-reporting/gsc",
"GSC_ALLOW_DESTRUCTIVE": "false"
}
}
}
}
Use full paths, starting with /Users/. The Claude app does not read your Terminal's
PATH, so a bare analytics-mcp will fail silently. Don't use ~
in this file either; it is not always expanded, so write the whole path out.
Restart Claude and test
Quit Claude completely with Cmd Q (closing the window is not enough) and open it again. Then ask:
List my GA4 properties and my Search Console sites.
If both lists come back, you are connected. If one comes back empty or with an error, jump to where I got stuck and search for what you see.
Building the dashboard
In Claude Steps 10 to 13. No dashboard code to write.
Once the tools worked, I did not write any dashboard code myself. I told Claude which products I wanted, which GA4 property and Search Console site belonged to each, and what I wanted to see.
Map each product to its properties
I asked Claude to list everything my account could see. It found 26 GA4 properties and 8 Search Console sites, most of them old side projects. I picked the five products I actually report on and confirmed the mapping. One product had a Search Console site but no GA4 property, so its tab shows search data only.
The Overview tab
This is the Monday-morning view, all products against the previous 28 days. Click the tabs to see how the dashboard is laid out.
Product scorecard
Daily impressions and weekly users, one row per product
Search Console
- Clicks, impressions, CTR, position, each with change
- Daily chart, 8 weeks
- Quick wins: positions 4 to 20
- Top 30 queries
- Pages in search
- Device and country
Google Analytics 4
- Users, sessions, engagement rate
- Weekly users and sessions
- Source / medium, top 10
- Landing pages
- Channels, devices, countries
- Events
- Four headline numbers: organic clicks, search impressions, GA4 users, GA4 sessions, each with its change
- A product scorecard table: clicks, impressions, CTR, average position, users, sessions and engagement rate per product
- Small charts of daily search impressions and weekly users, one row per product, each on its own scale
One tab per product
Each product gets its own tab with the detail I used to dig for in two tools (the Product A tab in the preview above). Search Console on one side, GA4 on the other.
The Quick wins table is my favorite part. Queries sitting between positions 4 and 20 are the cheapest traffic you will ever get, because a content update or a few internal links can move them onto page one or into the top three.
Make it refresh itself
I set up a scheduled task in Claude that runs every Monday morning before I start work. It pulls the last 28 days from both tools, updates the dashboard, and sends me a short summary: what moved, which product dropped, and any new quick-win queries.
Mac on, Claude open. Fresh numbers and a short summary before I start work.
Mac asleep. The task skips that week and tells me why.
The one catch
The MCP servers run on my Mac, so the Mac has to be on with the Claude app open when the task runs. If that is a problem for you, put a Google Sheet in the middle that pulls the data on its own schedule.
Where I got stuck, and how to fix it
Most of these happened to me, and the rest are the ones I checked for so they would not. Type a word from your error to find its fix.
The Clients page under Google Auth Platform is emptyStep 3
Why it happensNo OAuth client exists yet.
FixFinish Get started (Step 3) first, then + Create client.
Publish app asks for a website, privacy policy and moreStep 5
Why it happensScopes are listed on the Data Access page, which triggers verification.
FixDelete the scopes on Data Access, save, publish again.
Yellow banner: "Your app requires verification"Step 5
Why it happensNormal for a published, unverified app.
FixIgnore it. It works for up to 100 users.
"Google hasn't verified this app" at loginStep 7
Why it happensSame reason. This is your own app.
FixClick Advanced → Go to SEO Reporting (or whatever you named your app).
Login stops working after a weekStep 5
Why it happensThe app is still in Testing mode, which gives 7-day tokens.
FixPublish the app (Step 5), then sign in again.
Tools connect, but GA4 and Search Console return 0 propertiesStep 7
Why it happensYou signed in with a Google account that has no access.
FixRun the login again and pick the right account, or add the account you used as Viewer in GA4 and Restricted user in Search Console.
GA4 error 403 USER_PROJECT_DENIED: "Caller does not have required permission to use project"Step 8
Why it happensThe credentials bill API usage to a Cloud project (quota_project_id in the credentials file, or GOOGLE_PROJECT_ID / GOOGLE_CLOUD_PROJECT in the config), but the Google account you signed in with is not a member of that project.
FixRemove quota_project_id from the credentials file and the project variable from the config, then restart Claude. Or grant that account the Service Usage Consumer role.
Search Console works, GA4 says the API is disabledStep 2
Why it happensOnly "Google Analytics API" was enabled.
FixEnable Google Analytics Admin API and Google Analytics Data API.
Tools never appear in Claude, or spawn ENOENTStep 8
Why it happensClaude cannot find the program. Desktop apps do not read your shell PATH.
FixUse full paths in the config (Step 8).
Search Console server crashes on startStep 6
Why it happensIt needs Python 3.11 or newer.
FixInstall with uv tool install --python 3.12.
Changes do not show upStep 9
Why it happensClaude only reads the config at launch.
FixQuit with Cmd+Q, not just the window.
No match. Try one word from the error, like 403 or ENOENT.
Why is my top country missing from the Search Console data?
This one was subtle. One product's top country, the United States with 52 impressions, was missing from my country list.
The Search Console API sorts rows by clicks. Rows with zero clicks are not sorted by impressions after that; in my data they came back in what looked like alphabetical order. So when I asked for the top 8 rows, a country with 52 impressions and 0 clicks lost to "Argentina" with 2.
Top 8 countries
- India 14 clicks 310 impr.
- United Kingdom 3 clicks 61 impr.
- Germany 1 click 40 impr.
- Argentina 0 clicks 2 impr.
- Australia 0 clicks 9 impr.
- Brazil 0 clicks 5 impr.
- Canada 0 clicks 11 impr.
- France 0 clicks 4 impr.
- Japan 0 clicks 6 impr.
- Singapore 0 clicks 3 impr.
- United States 0 clicks 52 impr.
The fix: request as many rows as the tool allows, then sort by impressions yourself. The same applies to queries and pages. On small products this matters a lot, because most rows have zero clicks.
Is all this referral traffic spam?
Two of my products showed a lot of Referral traffic, mostly from Japan and Singapore. I assumed bots. The source / medium report told the real story:
taafttaaft_featFrom There's An AI For That, a real AI-tools directory.
accounts.google.comcheckout.stripe.comPeople coming back from Google sign-in and Stripe checkout.
To fix the second one, go to GA4 Admin → Data streams → your stream → Configure tag settings → List unwanted referrals and add both domains.
What the dashboard showed me in week one
The first run paid for the evening. Three things I had not spotted by clicking around GA4 and Search Console:
83%
of all search impressions came from one product. The other four together were barely visible. That changed where I spend link-building time.
8.9
was the position of a head term with 899 impressions and only 4 clicks in 28 days. Moving it into the top 3 is worth more than any new page I could write this month.
5.9
was where a product ranked for its own brand query, with 92 impressions and zero clicks. Similarly named sites likely take those clicks. That is a brand and title problem, not a content one.
None of these needed new data. They needed all the data in one place.
What it costs and what to watch
- Cost
- $0. The GA4 and Search Console APIs are free and only have daily limits, which a weekly report never comes close to. The Cloud project does not even need a billing account.
- Access
- Analytics access is read-only. Search Console access is technically full, but the add and delete tools stay off unless you turn them on.
- Secrets
- Your Google login lives in
~/.config/seo-reportingon your own computer. The dashboard only ever gets finished numbers. - Limit
- The dashboard is a weekly snapshot, not live. If you need history beyond 8 weeks, or the refresh cannot depend on your laptop, add a Google Sheet in the middle.
What to do this week
Tick things off as you go; your progress stays in this browser.
Rankings are a means, not the goal. A dashboard that puts every product on one page makes it obvious which one deserves this week's effort.