Search for "marketing MCP servers" and you will find hundreds of listings, many of them weekend projects that ask you to paste an API token into a config file. Some are excellent. Many are unmaintained, and a few would give an AI assistant write access to a live ad account with no guardrails. For a marketing or RevOps team, the sensible starting point is narrower: the servers published by the platforms you already pay for. This list sticks to servers I could verify from the vendor's own documentation or repository, as of October 2026, and notes where a detail comes only from third-party coverage. Status changes quickly in this space, betas become generally available and endpoints move, so treat every line below as something to re-check against the linked source before you connect production data.
CRM: HubSpot and Salesforce
HubSpot documents two MCP servers on its developer site. The remote server connects an MCP client to a specific HubSpot account so the assistant can read CRM data such as contacts, companies and deals. Setup runs through the HubSpot Developer Platform: you create a user-level app with read scopes for the objects you want to expose, then connect to the mcp.hubspot.com endpoint with OAuth. The second, the Developer MCP server, is a local, CLI-based tool for building HubSpot apps and is not what a marketing team needs. Salesforce offers Hosted MCP Servers, documented on developer.salesforce.com and labelled Beta in its official docs and its forcedotcom/mcp-hosted repository at the time I checked. The point of note for RevOps is per-user authentication: agent actions respect each user's existing Salesforce permissions. Salesforce also documents how Apex methods and REST resources can be surfaced as MCP tools, which means your admin team decides what an agent can and cannot do in the org. One third-party article says the hosted servers became generally available in April 2026. I could not confirm that from Salesforce's own pages, so check the current status before planning around it.
Paid media: Google Ads and Meta
Google maintains an open-source Google Ads MCP server in the googleads organization on GitHub, with documentation on the Google Ads API developer site. Google's docs describe it as read-only in its current release. It runs locally in Python, authenticates with OAuth or a service account, and needs a Google Ads developer token. Its tools are deliberately few: list the accounts you can access, and run Google Ads Query Language searches for metrics, budgets and status. That makes it a strong reporting and diagnosis tool and, by design, not a bid management tool. Meta's offering is newer. Trade coverage reports that Meta launched Ads AI Connectors in open beta on 29 April 2026: a hosted MCP server at mcp.facebook.com/ads plus a companion CLI, signed in through Facebook Business, with both reporting and write capabilities such as creating and editing campaigns. I could only verify these details through third-party reporting, not a Meta-hosted page, so read Meta's own documentation before connecting it. The write access is the part to plan for. I have written separately about what Meta Ads tool calls look like in practice and where they need human review.
Analytics and lifecycle: GA4 and Klaviyo
Google Analytics has an official MCP server in the googleanalytics organization on GitHub. It runs locally, is labelled experimental on its repository, and uses Google application default credentials with the read-only analytics scope. Its tools cover account and property details and running reports against GA4 data. Be careful with the name: a separate community project with a very similar name is published on PyPI, and it is not Google's. For email and SMS, Klaviyo documents a hosted MCP server at mcp.klaviyo.com/mcp, with setup guides for Claude, ChatGPT and Cursor. You authenticate with your Klaviyo account, and the docs restrict the remote server to users with the Owner, Admin or Manager role. Example uses in Klaviyo's docs are questions like how campaigns performed over the last 30 days or which flows convert best. Earlier versions of the guide labelled it beta. The current version does not, but as with everything here, confirm before relying on it.
How I evaluate an MCP server before connecting it
I run every server through the same six questions. Who publishes it? A server from the platform's own GitHub organization or developer docs is the default choice. A community server can be fine, but read its code and check when it was last updated. What can it write? Read-only servers are low risk. A server that can change budgets, pause campaigns or edit CRM records needs the same controls you would apply to a new junior hire: limited scope, a review step and a log. How does it authenticate? OAuth with scoped permissions is far better than pasting a long-lived token into a config file, which then sits on a laptop. Where does it run? A hosted server keeps credentials off your machine. A local one keeps data off a third party's. Both are valid, but know which you have. What does it log, and where? You want to know which tools the assistant called and with what effect. And is it listed in the official MCP Registry? The registry verifies the publisher's namespace, by GitHub or by domain, though not the quality of the code, so it is a useful check rather than a guarantee.
Where calculators and research servers fit
Platform servers give an assistant access to your data. A second category gives it a method: servers that compute something or supply reference material. My own server is an example of that category, and I mention it here for completeness rather than as a recommendation over anything above. It is listed in the official MCP Registry as co.rahuldsarker/growth-tools and exposes 116 calculators for marketing, RevOps and SaaS unit economics, plus search over my articles and case studies. It needs no account because it touches none of your systems. The useful pattern is to combine the two categories. Pull spend and revenue from a platform server, then run a payback or LTV to CAC calculation on the result with a deterministic tool rather than letting the model do arithmetic from memory. Whichever servers you use, keep the number of connected servers small. Every additional server adds tools the model has to choose between, and some clients struggle with long tool lists.
A sensible rollout for a marketing team
Start read-only. Connect the GA4 and Google Ads servers, or your CRM's server with read scopes only, and use them for a few weeks of reporting questions: why did CPA move, which campaigns drove pipeline, which deals stalled. Compare the answers against your dashboards. If they disagree, find out why before trusting either. Only then consider write access, one platform at a time, with a rule that any change to spend or live campaigns is proposed by the assistant and approved by a person. Document which servers are connected, who connected them and with what scopes, the same way you would track any integration with access to customer or financial data. If your CRM holds personal data on Indian or US customers, involve whoever owns privacy before connecting an AI client to it. None of this is complicated, but the teams that skip it tend to find out what their assistant can do the hard way.
Sources
HubSpot, MCP servers: https://developers.hubspot.com/mcp . Google Ads API, MCP server: https://developers.google.com/google-ads/api/docs/developer-toolkit/mcp-server . Google Ads MCP repository: https://github.com/googleads/google-ads-mcp . Google Analytics MCP repository: https://github.com/googleanalytics/google-analytics-mcp . Salesforce Hosted MCP Servers: https://developer.salesforce.com/docs/platform/hosted-mcp-servers/guide . Salesforce mcp-hosted repository: https://github.com/forcedotcom/mcp-hosted . Klaviyo MCP server: https://developers.klaviyo.com/en/docs/klaviyo_mcp_server . PPC Land, Meta opens its ad system to Claude and ChatGPT with new AI connectors: https://ppc.land/meta-opens-its-ad-system-to-claude-and-chatgpt-with-new-ai-connectors/ . Model Context Protocol, Introducing the MCP Registry: https://blog.modelcontextprotocol.io/posts/2025-09-08-mcp-registry-preview/
FAQ
Yes. Google publishes an open-source Google Ads MCP server in the googleads GitHub organization, documented on the Google Ads API developer site. Google describes the current release as read-only: it lists accessible accounts and runs Google Ads Query Language searches. It runs locally and needs a developer token and OAuth or service account credentials.
Not at the start. Begin with read-only servers for reporting, check the answers against your dashboards, then add write access one platform at a time. Any change to spend or live campaigns should be proposed by the assistant and approved by a person, with a log of what changed.
Some are well built, but they vary widely. Read the code, check when it was last maintained, prefer OAuth over pasted API tokens, and know whether it can write to your accounts. A listing in the official MCP Registry verifies the publisher's namespace, not the quality or safety of the code.