ArticleAI Workflows

MCP Tools for Marketers: Servers for Ads, Analytics, Search Console and Keyword Research

A neutral map of MCP servers a marketer can use: ads, analytics, Search Console, keyword research, SEO and CRM. Who publishes each, what it can change, and how to vet one.

Published 18 min read
02 / In this article

MCP tools are servers that let an AI assistant read data from, or act in, a marketing platform through one open standard. For a marketer the useful ones sit in five places: advertising, analytics, Search Console, keyword research, and SEO or CRM suites. Some are published by the platform itself, others by independent developers, and the difference matters more than the feature list. This guide maps what we could verify in October 2026, with no ranking, and ends with a way to vet any server before you connect it.

Key Takeaways

  • An MCP server is a connector, not a product. The same standard links an assistant to Google Ads, to a CRM or to a keyword database, so you evaluate each server on who runs it and what it is allowed to do.
  • Official does not mean read-only, and community does not mean unsafe. Google's Ads server is documented as strictly read-only, while HubSpot's and Meta's hosted servers can create and edit. Check the tool list, not the label.
  • Search Console and keyword research are mostly community territory. In the places we checked we found no official server for either, so those servers need the most checking.
  • Safety comes from permissions you control. Read-only credentials, a separate account for the connector, spend and usage caps, and a human approving any change are what protect you, whoever wrote the server.
  • Everything here dates fast. Versions, plans and availability below are as of October 2026 and each claim links to the page where we read it.

What are MCP tools, and why do marketers care?

The Model Context Protocol (MCP) is an open-source standard for connecting AI applications to external systems, and an MCP server is the piece that exposes one system's data and actions to the assistant. The official introduction compares it to a USB-C port for AI applications: one standard connector instead of a custom integration per tool.

For marketing work this changes one thing in practice. Instead of exporting a report from an ad account, pasting it into a chat and asking for analysis, the assistant can query the account itself, within the permissions you grant. If the idea is new to you, our explainer on what MCP is and how it works covers the protocol itself. This article stays on the practical question: which servers exist for the platforms a marketing team already uses, and which ones deserve trust.

A server can be small, a handful of read-only tools, or broad, with tools that create and edit records. The tool list is the real description of what it can do.

How should you read the categories below?

Every entry below answers the same five questions, so you can compare without a ranking: who publishes it, what its tools can do (read or write), how it authenticates, how mature it looks, and where we read the facts.

  • Publisher and scope. "Official" means the platform's own documentation or organization publishes the server; "community" means an independent developer does. We call a server read-only only when its documentation says so.
  • Authentication and maturity. How it signs in, the labels the publisher uses, release tags, and the last code change we saw on GitHub on 22 October 2026. Recent activity is a signal, not a guarantee.

Where we could not confirm something from a primary page, we say so instead of filling the gap. This is a map of what exists, not advice to install any of it. Versai does not claim to use or recommend any specific server here.

Community, can writeRead every tool firstPlatform, can writeHubSpot, Meta AdsCommunity, read toolsA Search Console serverPlatform, mostly readGoogle Ads, AnalyticsWHO PUBLISHES IT →WHAT THE TOOLS CAN DO →
Two questions place a server before you read a feature list. The names are examples from this guide.

Which MCP tools exist for advertising?

For advertising, Google and Meta each publish a server, and they differ most on whether the assistant can change the account. Google's is read-only by design, Meta's lists tools that create and edit.

Is there an official Google Ads MCP server?

Yes. Google's Ads API documentation has a page for an MCP server, and the code lives at googleads/google-ads-mcp. The documentation states: "This implementation is strictly read-only." It cannot change bids, pause campaigns or create assets. It is written in Python and runs over standard input and output or over Streamable HTTP, for example on Cloud Run.

The README lists three tools. list_accessible_customers returns the accounts the signed-in user can reach, search retrieves account information, and get_resource_metadata describes a resource type such as a campaign and its fields. The documentation says the search tool runs queries in Google's query language for metrics, budgets and status. Authentication is OAuth 2.0 or a service account, and the README also describes a Google Cloud deployment that handles OAuth for you. Google's page, last updated 16 October 2026 by its own footer, also says developer tokens were sunset on 9 September 2026, which changes older setup guides. Hosted endpoint access is limited to allowlisted participants in a pilot, so you run it yourself.

On maturity: the repository is Apache-2.0 licensed, shows no tagged releases, and its last code push was on 9 October 2026. The README warns that the server "will expose your data to the Agent or LLM that you connect to it." Our dedicated walk-through is in the Google Ads MCP guide.

Is there an official Meta Ads MCP server?

Yes. Meta's developer documentation describes an Ads MCP server hosted by Meta at mcp.facebook.com/ads. Its overview lists seven tool categories: reporting, ad creation and management, catalog creation and management, signals and datasets, help and troubleshooting, A/B tests and conversion lift studies, and activity logs. The same page says the tools can "create and edit ads, ad sets, and campaigns."

That is the opposite scope from Google's server. The overview page we read does not describe authentication, does not give a status label and does not say whether write actions need confirmation. It points to a separate "Get started" page for requirements and sign-in, which we did not use for any claim here. If you test it, read that page first and start with the reporting tools only.

What about LinkedIn Ads, TikTok Ads and others?

We did not verify an official server from these vendors' own documentation, so we list none. Directory listings are written by each server's author and are not evidence; treat such servers as community software. Our paid media service covers the channels themselves.

Which MCP tools exist for analytics?

For analytics, Google publishes an experimental server for Google Analytics that is read-only by default. Other analytics tools may have servers too, but this is the only one we verified from its own repository.

Is there an official Google Analytics MCP server?

Yes, and it carries a label. The repository googleanalytics/google-analytics-mcp is titled "Google Analytics MCP Server (Experimental)", and Google's developer site links to it from a banner. It is a local server that uses the Analytics Admin and Data APIs.

The README lists these tools: get_account_summaries, get_property_details, list_google_ads_links, list_property_annotations, run_report, run_funnel_report, get_custom_dimensions_and_metrics and run_realtime_report. All are read-only, with one exception: create_property_annotation writes data, and it is only available when the server is started with an --allow-write flag. Authentication uses Application Default Credentials with the analytics.readonly scope, plus analytics.edit for the write tool.

That design is worth copying in any evaluation: reads on by default, writes behind an explicit switch and a separate permission. The repository is Apache-2.0, has no tagged releases, and its last code push was on 21 October 2026, the day before we checked. The PyPI package is called analytics-mcp. A full setup is in our Google Analytics MCP guide.

Connecting an assistant to your analytics?

We help scope what an assistant may read, set up read-only access and keep your tracking clean.

See analytics and tracking

An assistant reads whatever your property collected, so missing or duplicated events produce fluent summaries of a mess. Fix tracking first, which is the point of analytics and tracking work.

GOOGLE ADSANALYTICSHUBSPOTPublisherGoogleGoogle org repoHubSpotWritesStrictlyread-onlyRead, 1 writeflagRead and writeSign-inOAuth or serviceacctDefaultcredentialsOAuth with PKCERuns onYour machine orcloudYour machineHubSpot serversLabelHosted: pilotonlyExperimentalSome tools inbeta
Three servers, three different shapes. Facts as published in October 2026.

Which MCP tools exist for Search Console?

For Search Console we found community servers only. In the sources we checked, including Google's own developer pages, we did not find an official one, so each server needs the full vetting below.

What does a community Search Console server look like?

One widely listed example is ahonn/mcp-server-gsc, MIT licensed. Its README describes a single tool, search_analytics, which returns search performance data: queries, pages, countries, devices and search appearance, with filters and up to 25,000 rows per request. It also offers an opt-in feature that flags pages ranking in a position range you set with enough impressions.

The README documents service account authentication: you create a service account in Google Cloud, download a JSON key and add the account to your Search Console property. The page does not say the server is read-only, but the only tool it documents retrieves data. The GitHub API shows 281 stars and a last push on 4 September 2026. Treat that as a data point, not a verdict: stars measure attention, not review.

The service account is the part to plan around. It becomes a user of your property, so give it the lowest permission Search Console allows, use it for this connector only, and remove it when you stop. Our Search Console MCP guide goes through setup and the questions worth asking it.

Search Console data answers what changed, not why. Causes sit in content, technical state and competitors, which is where SEO, GEO and AEO work starts.

Which MCP tools exist for keyword research?

Keyword research is where official and community lines blur, because the data comes from an API that Google owns but its own MCP server does not expose. Community servers fill that gap, and a few SEO suites publish their own.

Is there a Keyword Planner MCP server?

Keyword Planner data is available through the Google Ads API, and several independent servers wrap it. The official Google Ads server we described lists only three tools, none of which generates keyword ideas, so we count Keyword Planner access as community territory in October 2026.

One example is ncosentino/google-keyword-planner-mcp, MIT licensed, last pushed on 11 October 2026. Its README lists four tools: generate_keyword_ideas, get_historical_metrics, get_keyword_forecast and search_geo_targets. It needs a Google Ads account with billing configured (no ads have to run), a Google Cloud project with the Google Ads API enabled, an OAuth client and a refresh token. It warns that without active ad spend, search volumes appear as ranges, and that its HTTP host does not authenticate callers, so keep it on loopback or behind an authenticated proxy. The README does not say whether it is read-only.

Which SEO suites publish their own servers?

Two do, and both are hosted by the vendor.

Ahrefs. Ahrefs documents a remote MCP server on paid plans starting from Lite, at https://api.ahrefs.com/mcp/mcp over Streamable HTTP. Connecting shows a consent screen, and approval creates an API key tagged MCP; you can also generate one manually. Keys created through OAuth for MCP default to a monthly limit of 200,000 API units per key, which admins can change. The documentation states: "The external MCP server is not a general-purpose programmatic API." Scripts and standalone clients are not allowed, and programmatic use goes through Ahrefs' public API. The page we read does not say whether the connection is read-only.

Semrush. Semrush's developer documentation lists an endpoint at https://mcp.semrush.com/v2/mcp, Streamable HTTP only. OAuth is the default, with an API key header as a fallback for clients that do not support it. It covers the SEO API, the Trends API and the read-only methods of the Projects API, so you can read project data but "not create new projects, modify settings, or perform other actions." Usage draws on the same API unit system as the Semrush API, and eligibility depends on your plan. The documentation page we read was long and we read the first part of it, so confirm plan details on Semrush's page before relying on them.

Neither tool replaces the other in this guide, and nothing here ranks them. The point is that the access model differs: a metered key you can cap, versus API units shared with everything else on the account.

Which MCP tools exist for CRM, email and social?

For CRM, HubSpot publishes a hosted server with read and write access. For email and social platforms, we did not verify official servers from the vendors' own pages, so we do not list any.

What does HubSpot's MCP server do?

HubSpot documents a remote server at https://mcp.hubspot.com. Its developer page lists read and write access to contacts, companies, deals, tickets, products, orders, quotes, subscriptions and lists, plus engagements such as calls, emails, meetings, notes and tasks. It lists read-only access to users, teams, owners, campaign metrics, landing pages, website pages and blog posts. Authentication is OAuth, and HubSpot's integration guide says clients must use OAuth with PKCE. Actions follow the signed-in user's HubSpot permissions, so the assistant sees no more than that person can. The server does not allow access to custom Sensitive Data Properties.

Some features carry a beta label, such as reading revenue objects, and some tools are limited to named AI clients. HubSpot also publishes a separate local developer server for app and CMS development, which is a different thing and not a marketing data connector.

A CRM server is the most sensitive one here: contacts hold personal data, and an assistant that can edit them can overwrite them. Our AI workflows work starts by deciding what an automation may touch.

Want an assistant wired to your marketing data?

Tell us which platforms you use. We scope read-only connections first and a review step for anything that changes.

Plan your AI workflow

How do you evaluate an MCP server before installing it?

Evaluate a server as you would any software that runs with your credentials: confirm who wrote it, read what each tool does, narrow the permissions, test it on a copy, and keep a way to turn it off. The MCP project's own security guidance is clear that local servers can run with the same privileges as the client and that scope design shapes the damage from a leaked token.

01IdentifyWhopublishesit, andwhere isthe code02ReadtoolsListeverytool andmark thewrites03NarrowRead-onlyscope,separatecredentials04TestRun iton asandboxor acopy05RevokeKnow howto cutaccessandrotate
A five-step check to run before any server touches a live account.

Step 1: Who publishes it, and is the code open?

Prefer a server in the platform's own documentation or organization, then check the repository: license, issues, last commit and tagged releases. Two Google servers above show no tagged releases, so pin to a commit you reviewed.

Step 2: What does each tool do?

Open the tool list and mark every tool that creates, edits, deletes or sends. Those are your risk. The Analytics server is a good model because its one write tool is off unless you start it with a flag. If a server mixes reads and writes with no switch, assume the assistant can use both.

Step 3: How narrow can you make the permissions?

Use a dedicated credential for the connector, never your personal admin login. Grant the lowest scope that does the job: for example the read-only Analytics scope, or a Search Console user with restricted access. The MCP security guidance calls this scope minimization and warns against broad or "omnibus" scopes, which widen the damage if a token leaks. Where the platform has a usage cap, such as the per-key unit limit in Ahrefs, set it low.

Step 4: Where does it run, and what can reach it?

A local server runs with your privileges. The MCP guidance recommends the stdio transport for local servers and restricted access if you use HTTP. Hosted servers move that risk to the vendor, in exchange for your queries passing through their infrastructure.

Step 5: Can you see what it did and switch it off?

Want logs of which tool ran, a way to revoke the credential in the platform's settings and a place to rotate keys. For anything that spends money or changes tracking, a human reviews the change first.

What can go wrong with an MCP server?

The main risks are a server that does more than intended, a credential broader than needed, an exposed endpoint, and data sent to an assistant you did not mean to share it with. Each has a control you own: read-only first, minimal scopes, a reviewed and pinned version, loopback or authenticated access, and a decision about which data an assistant may see. The Google Ads README says the server "will expose your data to the Agent or LLM that you connect to it."

How do these servers fit into a marketing workflow?

They fit best as the read layer of a workflow you already understand, with a person approving anything that changes. A typical pattern starts with read-only access to ads, analytics and Search Console, asks the assistant for a summary or an anomaly check, and ends with a human deciding what to do. Writes, like pausing a campaign or editing a CRM record, come later and pass through an approval step.

This is also where automation platforms and custom builds enter. You can connect a vendor's server as it is, or build a narrow server of your own that exposes only the two or three actions a team needs, as we explain in how to build an MCP server. A narrow server of your own is often safer than a broad one you did not write, because the missing tools cannot be misused.

Run a first-month trial: compare what the assistant reports with the platform's own interface and find out why any numbers differ.

Frequently Asked Questions

What are the best MCP tools for marketers?

There is no single best set. The right choice depends on which platforms you use and how much access you will grant. Start from the platforms where you already have clean data, pick the officially published server where one exists, and compare scopes before features.

Are official MCP servers safer than community ones?

Not automatically. An official server comes from the platform's own organization, which helps with accountability, but it can still hold write tools. A community server can be narrow and read-only. Judge each by its tool list, permissions and code, and use the checklist above.

Is there an official MCP server for Google Search Console?

We did not find one in the documentation and repositories we checked as of October 2026. Community servers exist, such as one that exposes a single search analytics tool. Check Google's developer pages again before you decide, because this can change quickly.

Can an MCP server change my ad campaigns?

It depends on the server. Google's Ads server is documented as strictly read-only. Meta's Ads server lists tools that create and edit ads, ad sets and campaigns. Check the tool list and keep a person in the approval path for any change.

Do I need a developer to use these?

Hosted servers from Ahrefs, Semrush and HubSpot are mostly sign-in and consent in a supported client. Local servers such as the Google ones need Python, Google Cloud credentials and a config file.

Does MCP work with ChatGPT as well as Claude?

MCP is an open standard, and its documentation lists both Claude and ChatGPT among supported assistants. Client support for particular servers and features varies, so check your client's documentation. We compare the two clients in our guides on Claude and MCP and ChatGPT and MCP.

Next steps

Start with the platform whose data you trust most and connect one read-only server to it. Follow our walk-throughs for Google Ads, Google Analytics and Search Console if those are your platforms, and read what MCP is if the protocol is still new to you.

If you would rather scope this with us, our AI workflows team can map which connections make sense, and the paid media, SEO, GEO and AEO and analytics and tracking services cover the work that feeds them. To talk it through, get in touch.

Sources

All pages read in October 2026. Server details change often.

Newsletter

One short email when we publish

Notes on AI, image, video and brand work, from the people who do it. No spam.

Author