Marketing Data MCP is a remote MCP server that lets ChatGPT and Claude query GA4, Search Console, Google Ads, Merchant Center, Meta and CallRail in plain language, through 72 tools and 125 expert presets.
MCP URLhttps://ga4mcpjs.vercel.app/mcp
Works with
ClaudeChatGPTany MCP client
6platforms
72tools
125expert presets
0databases
Pages near page oneChatGPT
ShareShareJ
J
Which of my pages are close to page one on Google?
One connector puts six marketing APIs behind the same conversation. Pick a platform to see the tools the assistant gets, what the platform is good at, and the limits the server states out loud.
72 toolsone server
Google Analytics 4
17 tools · 24 presets
Standard, realtime, pivot, funnel and cohort reports, batches, and one report across up to 10 properties. The Admin API adds key events, custom definitions, data streams, audiences, product links and change history.
run_ga4_preset
run_ga4_report
run_ga4_realtime_report
run_ga4_pivot_report
run_ga4_funnel_report
run_ga4_cohort_report
batch_run_ga4_reports
run_ga4_multi_property_report
get_ga4_metadata
check_ga4_compatibility
list_ga4_key_events
list_ga4_custom_definitions
list_ga4_data_streams
get_ga4_admin_resource
run_ga4_access_report
run_ga4_audience_export
list_ga4_properties
Good at
Channels, landing pages, ecommerce and event-driven key events.
Says out loud
Thresholding and sampling can suppress small segments.
Google Search Console
8 tools · 14 presets
Queries, pages, countries, devices and search appearance; striking-distance opportunities; Discover and Google News; period-over-period deltas; sitemaps and URL Inspection, across up to 10 sites.
run_search_console_preset
query_search_console
query_search_console_advanced
compare_search_console_periods
inspect_search_console_url
list_search_console_sites
list_search_console_sitemaps
get_search_console_sitemap
Good at
Authoritative organic query and page performance.
Says out loud
25,000 rows per request, and anonymised queries sum below the site total.
Google Ads
10 tools · 34 presets
Campaigns down to search terms and assets, Shopping and Performance Max, impression share and Quality Score, Keyword Planner and Reach Planner, and one GAQL query across every client of a manager account.
run_google_ads_preset
query_google_ads
search_stream_google_ads
run_google_ads_across_accounts
google_ads_keyword_planner
google_ads_reach_planner
search_google_ads_fields
get_google_ads_field
list_google_ads_accessible_customers
list_google_ads_customer_clients
Good at
Paid performance, search terms and conversion reporting.
Says out loud
Auction Insights is not in the API. One call is one request, never a silent quota drain.
Google Merchant Center
14 tools · 13 presets
Product status and item issues, free listing and Shopping performance by product, brand, category and country, price competitiveness and insights, best sellers and competitive visibility.
run_merchant_preset
search_merchant_reports
get_merchant_product_status_summary
get_merchant_account_issues
list_merchant_products
get_merchant_product
get_merchant_resource
describe_merchant_report_fields
list_merchant_accounts
get_merchant_account
list_merchant_subaccounts
list_merchant_data_sources
get_merchant_developer_registration
register_merchant_developer
Good at
Feed health and product-level visibility, with or without Google Ads.
Says out loud
The Cloud project must be registered once per account; that is the only write tool.
Meta
10 tools · 22 presets
Meta Ads insights with age, gender, country, platform, device and placement breakdowns; Facebook Page and Instagram organic insights; creatives, pixels and audiences; and competitor ads from the Ad Library. Connected as an optional second sign-in.
run_meta_preset
run_meta_insights
describe_meta_insights_fields
get_meta_assets
search_meta_ad_library
query_meta_graph
list_meta_ad_accounts
list_meta_pages
list_meta_instagram_accounts
get_meta_token_info
Good at
Delivery, spend and organic reach across Facebook and Instagram.
Says out loud
Meta's attribution windows differ, so compare trends with GA4, not totals.
CallRail
9 tools · 18 presets
Calls by source, campaign, keyword, landing page and tracker; answered versus missed; first-time versus repeat callers; and daily trends, normalized so calls join Google Ads, GA4 and Search Console rows. CallRail uses an API key set on the server, so it suits self-hosted deployments.
run_callrail_preset
list_callrail_calls
get_callrail_call
get_callrail_call_summary
get_callrail_call_timeseries
list_callrail_trackers
list_callrail_accounts
list_callrail_companies
get_callrail_resource
Good at
Phone-call attribution that joins the rest of the funnel.
Says out loud
Presets aggregate one page of calls; check fetched against total.
Plus four cross-platform tools: list_marketing_presetsget_marketing_schemalist_marketing_guardrailsnormalize_marketing_records
02 How it connects
From one URL to your first answer.
You paste one address. The client registers itself, you sign in with Google, and the server hands back an encrypted token. No API keys to copy and no database behind it.
STEP 01
Paste one URL
Add the MCP URL as a custom connector in Claude or ChatGPT, or with one command in Claude Code.
connector url https://ga4mcpjs.vercel.app/mcp
GET /mcp → 401 + resource metadata
STEP 02
The client registers itself
Dynamic client registration (RFC 7591) at /register returns a client_id. Redirect URIs must match what was registered.
POST /register { "client_id": "mcp_7f3c…", "token_endpoint_auth_method": "none" }
STEP 03
Sign in with Google
Authorization code with PKCE (S256). The account chooser always appears, so you pick which Google account the assistant can read.
choose an account
you@company.com
agency@partner.com
STEP 04
Add Meta, or skip
An optional second step for Facebook and Instagram. Skipping still leaves a fully working Google connection; nobody is sent to Facebook without choosing to.
Add Meta too?
Connect MetaSkip
both paths finish connected
STEP 05
A sealed token comes back
The server mints its own encrypted MCP token with the Google refresh token sealed inside. A serverless cold start never signs you out.
POST /oauth/token → access_token
issuer · resource · scope · expiry · sealed
STEP 06
Ask in plain language
The model picks a tool, the server calls the platform API, and you get rows, the request it made, notes and guardrails.
You don't need to know report names or metrics. Ask in your own words; behind each of these questions is a ready-made expert report the assistant can run. Pick a platform, or search, and click a question to copy it into Claude or ChatGPT.
125questions across six platforms
Google Analytics 4
Where visitors come from 6
Pages and content 3
Engagement and conversions 5
Online store 5
Who my visitors are 5
Google Search Console
Rankings 4
Who finds me, and how 5
Trends 3
Discover and News 2
Google Ads
Performance 7
Competition and quality 2
Shopping and Performance Max 4
Audience and targeting 7
Conversions and landing pages 3
Account setup 5
Video and calls 2
Audits 4
Google Merchant Center
Product performance 5
Feed health 1
Pricing 2
Market and competitors 5
Meta (Facebook and Instagram)
Facebook and Instagram ads 7
Who sees my ads 6
Facebook Page 4
Instagram 5
CallRail
Calls at a glance 3
What makes the phone ring 7
Who calls 4
Call quality 4
No question matches that. Try “revenue”, “mobile” or “calls”.
The numbersStraight from the platform, never estimated.
The report it ranSo anyone can check exactly what was asked.
Rows that line upGoogle Ads, Meta and GA4 side by side.
Honest notesWhat the data can and can't tell you.
04 Guardrails
It tells you what it can't do.
Every response carries platform guardrails, so the assistant states an API's limits instead of filling the gap with a confident guess.
You ask
Show me Auction Insights for my top campaign.
A guess
competitor-a.com · IS 41% competitor-b.com · IS 33% overlap rate · 27%
This server
Auction Insights isn't exposed by the Google Ads API, so no tool here can return it. Impression share is available instead, through the impression_share preset.
You ask
What was my keyword Quality Score last March?
A guess
Mar · QS 6 Apr · QS 7 May · QS 7
This server
Quality Score is a current attribute, not a history. The quality_score preset returns today's snapshot and says it ignored the date range.
You ask
Add up clicks from all my Search Console queries.
A guess
Σ query clicks = site total 48,210 clicks
This server
Search Console anonymises some queries, so query rows add up to less than the site total. Use the page or site total for that number.
You ask
Why don't Meta conversions match GA4?
A guess
tracking is broken fix your pixel
This server
Meta attributes conversions on its own windows, so the counts won't tie out with GA4 or Google Ads. Compare trends, not totals.
You ask
Pull every call we got last year, grouped by source.
A guess
google / cpc · 1,204 direct · 880 (all calls counted)
This server
Presets aggregate one page of calls and report callsFetched against totalRecords. If they differ, it pages rather than pretending the total is complete.
05 One schema
Six schemas in. One table out.
Every platform names things differently. Presets emit rows in one normalized schema, so an answer can compare Google Ads spend, Meta spend, GA4 sessions and CallRail calls side by side. get_marketing_schema documents it, and normalize_marketing_records maps anything else into it.
metrics.cost_microsGoogle Ads
spendMeta
sessionSourceMediumGA4
sourceCallRail
keys[0]Search Console
offer_idMerchant Center
cost
source_medium
source
query
product_id
platform
date
campaign_name
cost
clicks
conversions
google_ads
2026-09-21
Brand · Search
412.80
1,096
38
meta
2026-09-21
Retargeting · Reels
265.10
742
21
ga4
2026-09-21
Brand · Search
—
—
41
callrail
2026-09-21
Brand · Search
—
—
9
Rows are illustrative. Field names on the left are each platform's own.
06 Security
Stateless by design.
There is no database. Your connection lives inside an encrypted token that only this server can open, and every tool but one only reads.
OAuth with dynamic client registration and PKCE (S256); a redirect URI is accepted only if it was registered, or is a loopback address for native apps.
Read-only everywhere except register_merchant_developer, which registers the server's Cloud project on a Merchant Center account.
/mcp accepts only tokens this server issued, never raw Google tokens.
Meta requests are signed with appsecret_proof, Page tokens are never returned, and tokens are stripped from logs.
Quota-safe: one tool call is one API request, with no silent auto-pagination.
issuerhttps://ga4mcpjs.vercel.app
resourcehttps://ga4mcpjs.vercel.app/mcp
scopeanalytics.readonly · webmasters.readonly · …
expiryshort-lived · refreshable
google credentials████████████ sealed
MCP access tokenencrypted · no database
An illustration of what the encrypted token carries.
07 Setup
Connect in under a minute.
Add the server with this command. -s user makes it available in every project.
Open /mcp in Claude Code and choose Authenticate next to ga4mcp.
Sign in with Google, then connect Meta or skip it.
Ask: “Which landing pages lost the most organic clicks this month?”
terminal
$ claude mcp add --transport http -s user ga4mcp https://ga4mcpjs.vercel.app/mcpAdded HTTP MCP server ga4mcp> /mcp
ga4mcp · needs authentication → Authenticate
Open Settings → Connectors and choose Add custom connector.
Paste the MCP URL and give it a name such as “Marketing Data”.
Connect: sign in with Google, then connect Meta or skip it.
In a chat, enable the connector and ask your question.
connector url
Name Marketing Data
Remote MCP URLhttps://ga4mcpjs.vercel.app/mcpAuth OAuth · discovered automatically
Add a custom connector in ChatGPT's connector settings.
Paste the MCP URL. ChatGPT reads the OAuth settings from the server's discovery endpoints.
Sign in with Google, then connect Meta or skip it.
Use the connector in a chat and ask in plain language.
discovery
MCP URLhttps://ga4mcpjs.vercel.app/mcpmetadata /.well-known/oauth-protected-resource
auth server /.well-known/oauth-authorization-server
PKCE S256 · grant authorization_code, refresh_token
08 Reference
Marketing Data MCP at a glance
Key facts
What
A remote MCP server for marketing analytics
Endpoint
https://ga4mcpjs.vercel.app/mcp
Status
Public beta
Clients
Claude Code, Claude Desktop, claude.ai, ChatGPT and other MCP clients
Platforms
GA4, Search Console, Google Ads, Merchant Center, Meta, CallRail
Tools
72, plus optional Tag Manager audit and BigQuery GA4-export tools on self-hosted deployments
Marketing Data MCP is a remote MCP server built by Jayant Solanki. It lets ChatGPT, Claude and other MCP clients query Google Analytics 4, Google Search Console, Google Ads, Google Merchant Center, Meta and CallRail in plain language, through 72 tools and 125 expert presets.
What is an MCP server?
MCP, the Model Context Protocol, is an open standard for connecting AI assistants to tools and data. An MCP server publishes a list of tools; the assistant decides which tool to call, the server runs it against a real API and returns the result, and the assistant answers from that result.
Which AI clients work with it?
Claude Code, Claude Desktop, claude.ai and ChatGPT custom connectors, plus any MCP client that supports remote servers over HTTP with OAuth. Point the client at https://ga4mcpjs.vercel.app/mcp.
How do I connect it to Claude?
In Claude Code run: claude mcp add --transport http -s user ga4mcp https://ga4mcpjs.vercel.app/mcp, then open /mcp and choose Authenticate. In Claude Desktop or claude.ai go to Settings, Connectors, Add custom connector, and paste the same URL.
How do I connect it to ChatGPT?
Add https://ga4mcpjs.vercel.app/mcp as a custom connector in ChatGPT. ChatGPT discovers the server's OAuth settings automatically, registers itself and sends you to Google sign-in.
Does it change anything in my accounts?
Every tool reads data except one: register_merchant_developer, which registers the server's Google Cloud project on a Merchant Center account so the Merchant API can be used. Nothing is written to Analytics, Search Console, Google Ads, Meta or CallRail.
Is my data or login stored?
The server has no database. It issues its own encrypted MCP token with the Google refresh token sealed inside, so nothing about your connection is written to storage. Report results go back to your AI client and are not kept by the server.
What are presets?
Presets are ready-made expert reports: 34 for Google Ads, 24 for GA4, 22 for Meta, 18 for CallRail, 14 for Search Console and 13 for Merchant Center. Each returns the raw API output, the request it made, rows in a normalized cross-platform schema and the platform guardrails.
Why do Meta conversions not match GA4 or Google Ads?
Meta attributes conversions on its own attribution windows, so its counts will not tie out exactly with GA4 or Google Ads. The server says this in its guardrails and recommends comparing trends rather than absolute totals.
Can it query several properties or sites at once?
Yes. run_ga4_multi_property_report runs one report across up to 10 GA4 properties, query_search_console_advanced covers up to 10 Search Console sites, and run_google_ads_across_accounts runs one GAQL query across the clients of a manager account.
What does public beta mean here?
The server is live and in active development. Some platforms depend on outside approvals, such as Meta App Review and the Google Ads developer token's access level, so availability can vary by account and features may change.
Who built Marketing Data MCP?
It was designed and built by Jayant Solanki, an AI-ready SEO strategist and marketing technologist based in India.
Your data already has the answers. Let it talk.
Copy the URL, add it to Claude or ChatGPT, sign in once. Then ask the question you usually build a report for.
Built by Jayant SolankiAI-ready SEO strategist and marketing technologist. This server grew out of the reports I kept rebuilding for clients across GA4, Search Console and Google Ads. See the portfolio or try TagSpy.
MCP server for ChatGPT & Claude 
Claude
ChatGPT