MCP servers for marketing: 11 official options and how to choose your first connection

Last updated: 9 September 2026. Each vendor entry below was checked against its own documentation on this date.

A marketing team can connect five platforms and still fail to answer one basic question: which campaigns generated customers who stayed? The missing pieces are often a common identifier, an agreed definition of revenue and enough history to measure retention.

MCP makes systems accessible to an assistant. The value comes from giving that assistant a specific job with a verifiable result.

Choose a workflow before choosing a server

Start with a task you repeat and can already check manually.

For a demand-generation team, that might be a weekly account-level pipeline review in HubSpot. For a growth team, it might be diagnosing a signup funnel in Mixpanel. For an ecommerce team, it might be comparing Meta campaign performance against a trusted order model.

Write the task in this form:

Question: Which part of the trial funnel changed last week?
Source: Approved production analytics project
Definition: Existing saved funnel, same filters and conversion window
Output: Change by step, affected segments and report links
Actions: Reporting only
Reviewer: Growth operations

If the source or definition is missing, resolve it before adding more connections. Martechdevs' AI-ready data work focuses on that foundation: identities, usable models and definitions that support a real business question.

Open full diagram

CRM and paid media connections

1. HubSpot

HubSpot provides a remote MCP server and client-specific connector routes. Current documentation includes read and write capabilities, with access determined by permissions and the integration. Use the official client setup for the application you are connecting. HubSpot developer documentation.

A useful first job is an evidence-backed pipeline review: deals approaching their close date, missing next steps and owner follow-up requirements. Require record IDs and source links so the output can become a reviewed work list.

Delay bulk property updates until you have checked how your workflows react to those changes. Updating a seemingly harmless field can enrol contacts in automation.

The HubSpot MCP setup guide separates the connection paths and provides prompts for a controlled first workflow.

2. Google Analytics

Google's official Analytics MCP project is a local, experimental server for reading Analytics data. It uses Google credentials and Analytics APIs, rather than a universal hosted URL that you paste into any web application. Google Analytics MCP repository.

Start with a saved reporting question: sessions and key events by a defined acquisition dimension over two completed periods. Specify the property, timezone and exact metric names.

Keep user, session and event scope distinct. An assistant can confidently compare metrics that answer different questions unless the request makes their meaning clear.

The Google Analytics MCP guide covers local setup and alternatives for teams that need a hosted connection.

3. Google Ads

Google's official Ads MCP implementation currently supports read-only reporting. Its deployment options include a local process and self-hosting; account access follows Google Ads API requirements. Google Ads MCP documentation.

A strong first task is a campaign comparison that includes spend, the selected conversion actions and conversion value. Add the account currency and reporting dates to every result.

Do not rank lead-generation campaigns solely by cost per lead when sales acceptance or customer conversion varies substantially. Give the assistant the downstream outcome model, or have it state that the evidence stops at the lead.

The Google Ads MCP guide covers the current access setup and a bounded reporting query.

4. Meta Ads

Meta operates a hosted server at https://mcp.facebook.com/ads. Its documented capabilities include reporting and campaign management. Asset permissions and the controls available to the business determine what the connected agent can do. Meta's server overview.

Begin with a read-only performance review and a draft action list. Use account IDs, attribution settings and the selected conversion event in the report.

Campaign creation, creative editing and publication need separate review. Treat an attractive preview as one part of approval; budget, destination, audience and schedule also need checking.

The Meta Ads MCP guide explains the advertiser and developer setup paths, account rules and drafting workflow.

Warehouse and product analytics connections

5. BigQuery

Google provides a managed remote server at https://bigquery.googleapis.com/mcp. Its tools include metadata access and read-only or broader SQL execution. MCP Toolbox for Databases is a separate configurable option for teams operating their own server. BigQuery MCP documentation.

A useful first question combines data already joined in an approved reporting model, such as paid conversion by acquisition cohort. Avoid making every user session an exercise in rediscovering raw database relationships.

Query cost and result completeness deserve explicit tests. A short returned answer does not imply a small scan or a complete underlying result.

6. Snowflake

Snowflake supports managed MCP server objects that expose configured tools. The endpoint is specific to the account, database, schema and server name. Access also depends on privileges for the underlying objects. Snowflake-managed MCP documentation.

This is useful when the team has governed revenue or customer models and wants to make a narrow set of questions accessible. Start with a semantic definition that analysts already trust.

Inspect every tool added to the server. A server exposing a business report has a different risk profile from one exposing a procedure that changes data.

Our Snowflake MCP guide covers the connection in more depth.

7. Mixpanel

Mixpanel's hosted MCP server supports analysis, saved content and data-management tools. Its regional endpoints include US, EU and India. Some tool groups remain beta. Mixpanel MCP documentation.

Use it to investigate a defined funnel or retention question, then save the evidence in a report colleagues can reproduce. Event definitions and identity rules must be supplied alongside the question.

Do not treat this as a purely read-only connection: the current catalog contains operations that create or modify content and metadata. A pilot should explicitly limit which of those are available.

Mixpanel currently enables MCP by default on Free and Growth and disables it by default on Enterprise; administrators can change the setting. Check the actual workspace before connecting a client. Mixpanel AI access policy.

8. Amplitude

Amplitude's remote server exposes analytics, taxonomy and content tools with OAuth access. It supports reads and writes; its documentation distinguishes the account-data server from a separate public documentation server. The US endpoint is https://mcp.amplitude.com/mcp, with a separate EU endpoint. Amplitude MCP documentation.

A practical starting task is explaining a change in a saved product funnel while preserving the original filters and conversion window. Ask for alternative explanations that require additional evidence.

Editing a tracking plan or experiment configuration deserves a different permission and approval policy from reading a chart. Keep that distinction visible when granting access.

Enrichment and lifecycle connections

9. Clay

Clay's MCP supports find-and-enrich tools and enabled functions. Its platform connection endpoint is https://api.clay.com/v3/mcp. Administrators can manage access and credit limits for users. Clay connection documentation, MCP administration.

Choose a bounded research job, such as enriching a small reviewed account list with a clearly defined qualification criterion. Return the source and date for each factual finding.

Estimate credit use before scaling. An enrichment function that is affordable for 20 accounts can become expensive when an agent repeats it across several thousand records or retries without checking previous results.

Keep the list-generation decision separate from permission to contact the people found.

10. Customer.io

Customer.io's remote MCP server supports reading workspace information and managing lifecycle work. Its official endpoints include US and EU regions, and the documentation warns that connected tools can return information about people in the workspace. Customer.io MCP setup.

A useful first job is reviewing one lifecycle journey: entry conditions, exclusions, message sequence and exit behaviour. The result should identify specific campaign objects and a proposed correction list.

Before enabling sends or campaign changes, test suppression, audience counts and transactional versus marketing intent. A good message sent to the wrong population is still a bad outcome.

11. Salesforce Marketing Cloud Engagement

Salesforce has a generally available, hosted MCP server for Marketing Cloud Engagement. It exposes Engagement capabilities such as data extensions and journeys. Access combines the installed package's scopes with the user's permissions, and the endpoint is specific to the tenant and package. Salesforce announcement, setup documentation.

Begin by inspecting a journey's configuration or documenting the data extensions it depends on. Require a reviewed change plan before allowing the assistant to alter active marketing operations.

Salesforce has several MCP products. A developer tooling server, a CRM data server and the Marketing Cloud Engagement server are different choices. Confirm the exact product before following a connection tutorial.

Also relevant: Braze, with feature-specific availability

Braze documents an official MCP tool catalog containing reporting and modification operations. Several campaign, segment and Operator tools are restricted to their respective beta programs. Braze's available functions.

For an existing Braze team, that makes a reporting or campaign-inspection pilot worth evaluating. It does not establish that every account can create and launch campaigns through the server today.

Use the tools actually enabled in your workspace as the acceptance boundary. Plan limitations, beta enrolment and role permissions can all change the answer.

How to connect a remote server in Claude Code

For a remote HTTP server that supports the client's authentication flow, Claude Code uses a command in this form:

claude mcp add --transport http SERVER_NAME SERVER_URL

Then use /mcp to inspect and authenticate the connection. Some vendors need additional OAuth configuration or a different setup path. Local servers use a command and arguments instead of a remote URL. Claude Code MCP documentation.

Complete four checks before calling the setup finished:

1. The intended account or project is visible.
2. The tools match the approved workflow.
3. A known-answer query agrees with the source platform.
4. An operation outside the intended scope is denied.

Record the configuration location and owner. A connection installed for one project should not unexpectedly become available in every unrelated project a colleague opens.

How to evaluate a community server

An official server has a clear vendor relationship. That helps with provenance, but you still need to evaluate its permissions and behaviour. A community server needs the same checks plus a closer review of its maintainer and distribution.

Before installation, inspect the actual package, source repository, update history and requested access. Confirm that the endpoint you are using belongs to the provider you evaluated.

Review how it handles tokens, logs and outbound network calls. If you cannot explain where customer data goes, the evaluation is unfinished.

Pin the reviewed version when the workflow depends on locally installed code. Define who approves updates and how the acceptance questions will be rerun after a change.

The MCP security checklist turns this into a repeatable review. Repository stars and a polished demo are insufficient evidence of safe operation.

Worked example: three connections, one revenue question

Consider a fictional 18-person SaaS company using Google Ads, HubSpot and BigQuery. Its growth lead wants to identify campaigns that create customers with good early retention.

The team first connects Google Ads and HubSpot independently. It can retrieve campaign spend and closed-won deals, but it has no reliable mapping between campaign interactions and accounts. The assistant can produce two reports; it cannot yet produce the requested joined conclusion.

The next step is a warehouse model with account IDs, acquisition rules, payment dates and retention observations. The team validates a sample of records and exposes a cohort report through BigQuery.

Now the weekly workflow has a clear sequence: inspect ad spend, retrieve the approved cohort report, flag missing mappings, then draft questions for the growth meeting. Budget recommendations remain subject to review.

Suppose 70 of 500 acquired accounts lack the required channel mapping. The report should disclose the 14% unknown share and show whether it is concentrated in a particular period. Silently discarding those accounts would make the comparison look more complete than it is.

The lesson is practical: add a connection when it closes an evidence gap. If the gap is an undefined join, solve the join.

At Martechdevs, we build the data access, definitions and working instructions behind repeatable AI workflows. Start with the free martech AI audit, or explore an AI assistant built around your approved workflows.