Claude for marketing

Connecting Claude or ChatGPT to Google Ads and Meta with MCP, safely

MCP is how an AI assistant reaches your real accounts instead of the screenshots you paste in. It is the single biggest step up in usefulness, and the easiest way to make an expensive mistake. Here is what exists and the rules we use.

By Nathan, who built TheCMO and runs an agency on it. Written from what we use every day, September 2026.

What MCP is, in one paragraph

MCP (Model Context Protocol) is a standard way to give an assistant tools. An MCP server for Google Ads says, in effect, “here is a tool that runs a report”. Claude or ChatGPT decides when to call it and reads the answer. Both support MCP: Claude natively, ChatGPT through developer mode.

The official servers

  • Google Ads (github.com/googleads/google-ads-mcp): three tools. search runs Google Ads Query Language reports, get_resource_metadata describes what fields exist, and list_accessible_customers lists the accounts you can reach. It needs a developer token with at least Explorer access. Google’s own warning: it exposes your data to whatever assistant you connect.
  • Meta Ads: Meta launched an official server in April 2026 that can make changes as well as read. Reports say anything it creates starts paused and its write tools are off until you turn them on.

What a connection does not give you

Access to the reports is not knowing which ones to read. Our Google Ads audits pull 34 datasets because the expensive problems hide in the ones people skip: negatives at three levels, shared budgets, change history, policy status. A connected assistant with no method still reads the campaigns tab and guesses. How to give it the method.

The safety rules we use

  1. Start read-only. Weeks of reading before any write access. You learn what it gets wrong for free.
  2. Preview every change. Before anything live moves, capture the before-state, the exact new values, the spend at risk and how to undo it.
  3. A person approves. The assistant proposes. It never approves its own change.
  4. One attempt, then read it back. If the result is unclear, check the account. Never retry blindly: a retry applies the change twice.
  5. Keep a receipt. What changed, who approved it, what the account said afterwards. Next month you will want to know.
  6. Watch the change history. Google keeps 30 days and every entry says who made it. We found Google had auto-applied one of its own recommendations and removed two targeting settings that nobody approved.

A lesson we paid for

On our own system, 23 changes once reached a provider without their record being finished, because the process was stopped while it was reading the result back. The changes were real; the proof was missing. That is why the readback and the receipt are separate steps, and why an unclear result gets checked, never re-run.

Straight answers

What is an MCP server for marketing?
A connector that lets an AI assistant such as Claude or ChatGPT read from, and sometimes change, a tool like Google Ads or Meta Ads. The assistant calls it the way you would click a report.
Is there an official Google Ads MCP server?
Yes. Google publishes one on GitHub (googleads/google-ads-mcp). It has three tools: search, get_resource_metadata and list_accessible_customers. It needs a Google Ads developer token with at least Explorer access.
Can an MCP server change my ads?
Some can. Meta’s official ads server is reported to make changes, with anything it creates starting paused. Treat any write access as live spend and put an approval step in front of it.
Does MCP make Claude good at Google Ads?
No. It gives Claude the data. Good judgement still needs your strategy, what each number means and a fixed way to investigate.

Keep reading

Everything here is free to use. If you would rather have it built around your playbooks, installed and connected to your accounts, for your team to run or for us to run for you, that is what TheCMO is.

See TheCMO on a real agency