Google Ads MCP: how to connect an AI assistant to your ad account safely

By the AgentMsgr team · Updated November 11, 2026 · 10 min read

A Google Ads MCP is a connector that lets an AI assistant like Claude or ChatGPT use your Google Ads account. It can read campaigns and costs, pull keyword ideas and, with some connectors, build or change campaigns. There are three main routes this fall: Google's own server, which is free and built mostly for reports; hosted services like Pipeboard, which can build and change whole campaigns; and Zapier, which covers a short list of actions. Whichever you pick, let the assistant read freely and make it ask you before anything that spends money or changes a live ad.

MCP is short for Model Context Protocol, the shared standard these connectors follow. Everything below comes from Google's, Pipeboard's, Zapier's, Anthropic's and OpenAI's own pages; the dates each was read are in the sources at the end.

What is a Google Ads MCP, and what can it do?

Think of it as a plug between your assistant and your ad account. You sign in once, and the assistant can use the connector's tools, small actions like "run this report" or "add these negatives." Which jobs it can do depends on the connector:

• Reading: campaigns, ad groups, keywords, search terms, costs and conversions for any date range. On connectors that include Keyword Planner, also keyword ideas and historical search volumes.
• Changing: creating campaigns and ad groups, adding keywords and negative keywords, writing ads, pausing or turning things on, and editing bids and budgets. Only connectors with write access do this.

Reading is where most of the value is, and it's safe. A search terms review that takes an hour by hand takes a minute.

Changes are where it can go wrong. A wrong budget spends real money.

Is a Google Ads AI agent the same as AI inside Google Ads?

No. People looking for AI Google Ads help often mean Google's own automation, like Smart Bidding and Performance Max, which runs inside Google Ads. A Google Ads AI agent is your own assistant working from outside: it reads the account through a connector and, if you let it, makes the changes you'd make by hand. Many advertisers use both.

Google Ads MCP server options in November 2026

Four routes, side by side. Prices appear only where the company publishes them.

Google's official Google Ads MCP server

Google publishes an open-source server at googleads/google-ads-mcp on GitHub. Its guide lists seven tools. Three read: one lists your accounts, one runs reports in GAQL (Google's query language for ads data, so the assistant can ask for any metric over any dates), and one explains which fields a report can include. Four change assets and labels: create an asset, upload a video asset, manage labels, and attach labels to items. Google calls the mode "Read and partial mutate." Switch off the assets and labels tools in its settings file and it becomes read-only.

Those change tools came with version 0.0.6 on November 2, 2026, built as a propose-and-confirm flow: the change is proposed first and made only after a confirm step. The guide is plain about the limits. The server can't create, change, pause or turn on campaigns, and the same goes for ad groups, ads, budgets and bids. Keyword Planner ideas aren't among its tools either.

Setup is the hard part. You need a Google Cloud project with Google Ads API access (since developer tokens were retired on September 9, 2026, access comes from the project's level: Explorer, Basic or Standard), sign-in credentials, and Python with pipx. For Claude or ChatGPT on the web, you also deploy it to Google Cloud Run. It suits someone technical who mainly needs reports.

Pipeboard: a hosted Google Ads MCP with write access

Pipeboard is a hosted service, so it runs the connector for you and there's nothing to install. Its Google Ads guide, updated November 10, 2026, lists 70 tools. They include Keyword Planner ideas, historical metrics and forecasts, search term reports, and tools that create and edit campaigns, ad groups, ads, keywords and bids. New campaigns are created paused, so nothing spends until you switch it on. You can also start with a read-only token that can't change anything.

There's a free plan (prices are in the questions below), and the same login covers Meta, TikTok, LinkedIn and other ad platforms. The Meta Ads MCP guide covers the Meta side.

Zapier and open-source Google Ads MCP servers on GitHub

Zapier MCP gives your assistant one address for Zapier's app actions. For Google Ads the list is short: create a report, set a campaign's status, manage customer lists and send offline conversions. Each call uses 2 tasks from your Zapier plan, so it suits simple jobs like pausing a campaign from a chat.

Community servers on GitHub are the do-it-yourself route. A popular one, cohnen/mcp-google-ads, covers campaign data, metrics and keyword analysis, per its description; its code was last updated in October 2025. Check a project's last update before you hand it your account.

One rule to know about hosted services. Google's Developer Policies, updated August 31, 2026, ban "programmatic proxies": hosted services, MCP servers included, that only pass Google Ads access along as a middle layer. The same policy allows open-source tools you download and connect with your own access. If you pick a hosted connector, especially for a client's account, ask the vendor how it meets these rules.

How to connect a Google Ads MCP to Claude or ChatGPT

These steps use a hosted connector. For Google's own server, finish the Cloud Run setup first, then use its address the same way.

1. Pick the connector and copy its address. With Pipeboard, connect Google Ads on its Connections page and use https://google-ads.mcp.pipeboard.co/.
2. Choose the Google account that signs in. For reports and research, use one with Read only access in Google Ads. Google says that level can view campaigns and use planning tools but can't edit campaigns. Use Standard only when you want changes.
3. Add it to your assistant, using the steps below.
4. Set tool permissions before your first prompt.
5. Test with a read-only question, like "List my Google Ads accounts and last month's cost by campaign."

Claude Google Ads setup, step by step

Claude can add a custom connector on every plan, Free included, though Free allows only one. On Pro or Max:

1. Go to Customize, then Connectors.
2. Click + Add, then Add custom connector.
3. Name it (Google Ads works), paste the address and click Continue.
4. Review the sign-in settings Claude detects, click Add, and sign in when asked.
5. In a new chat, click + at the lower left, open Connectors and switch it on.

On Team and Enterprise plans, an Owner adds the connector first under Organization settings, then Connectors. Next, open the connector under Customize, then Connectors, and set each group of tools to Always allow, Needs approval or Blocked.

ChatGPT for Google Ads: who can add a custom connector

ChatGPT is stricter. When this was written (November 11), OpenAI's help center said full custom MCP connections, the kind that can make changes, are for Business, Enterprise and Edu workspaces on the web, created by admins and owners. Pro users can connect a custom server for reading only.

To add one, open Plugins, click the plus button and choose Add custom MCP server. Enter a name and the address, set up sign-in, tick the risk box and choose Create as a plugin. In a chat, type @ and pick it. ChatGPT asks you to confirm write actions, depending on the app's permissions.

What a Google Ads AI agent should do alone, and what needs your approval

A simple rule: reading is free, changing costs money. Let the agent read anything, and make every change come to you first, even the small ones.

Set it up in three places: the lowest Google Ads access level that does the job, write tools on Needs approval in your assistant, and the connector's own brakes, like Pipeboard's read-only tokens and paused campaigns. The guide to the human in the loop AI agent goes deeper on which actions should always wait for you.

Google Ads MCP permissions by goal

Match the setup to the job, and grant the least access that does it.

First prompts for AI PPC work: search terms, wasted spend and keyword research

AI PPC work goes best when your first prompts are read-only and specific. Swap in your details where you see brackets, and keep "make no changes" in each prompt until you trust the setup.

• Search terms review: "Pull the search terms report for [campaign] for the last 30 days. Sort the terms into relevant, unsure and irrelevant. For the irrelevant ones, show clicks, cost and conversions, and suggest a negative keyword with a match type for each. Make no changes."
• Wasted spend: "Find keywords and search terms that spent more than [$50] in the last 60 days with no conversions. Sort by cost. For each, say whether to pause it, lower the bid or add a negative, and why. Make no changes."
• Keyword research with historical volumes: "Get Keyword Planner ideas for [service] in [city or region]. For the best 30, pull 12 months of search volume and top-of-page bid ranges, and flag rising ones. Don't create anything." This needs a connector with Keyword Planner tools.
• Month over month: "Compare last month with the month before for each campaign: cost, clicks, conversions and cost per conversion. List the three biggest changes and a likely reason for each. Make no changes."
• A new campaign, drafted first: "Draft a search campaign for [service] in [city]: 2 ad groups, 10 phrase and exact match keywords each, 15 negatives and 2 ads per group, at [$X] a day. Show me the plan first. After I approve, create it paused."

When a review turns up fixes, approve one batch at a time: "Add negatives 1, 3 and 7 as phrase match. Show me the exact list first." To find good keywords in the first place, see how to find low competition keywords. To see what your ad clicks do once they land, a Google Analytics MCP covers that side.

Google Ads automation with two agents in one chat

Google Ads automation gets safer when a second agent checks the one that changes the account: an ads agent holds the connector, and a research agent questions its numbers before anything is applied.

AgentMsgr, which we make, is built for that kind of handoff between agents from different companies. A wasted-spend review could run there like this:

1. You post the goal: "Cut wasted spend on [campaign] without losing leads."
2. Your ads agent, the one with the Google Ads connector, pulls the search terms and wasted spend reports and posts what it found.
3. You @mention your Grok Bot agent, the one doing research, to check the suggested negatives against your list of services and flag any that would block good leads. SpaceXSI (formerly xAI) lists a paid media job like this among Grok Bot's uses: moving budget based on spend and cost per new customer, for your approval.
4. The ads agent asks for approval in the chat before it adds anything. You tap Approve or Decline, and only then does it make the change in Google Ads.

For the wider setup, see AI marketing team and the AgentMsgr MCP server guide. AgentMsgr never touches your ad account; changes happen through the agent's own connector. If that connector lives in your own Claude or ChatGPT, connect it to AgentMsgr and it can post and read in the chat for you.

Sources: Google, Google Ads MCP server guide and googleads/google-ads-mcp on GitHub (release 0.0.6, November 2, 2026); Google Ads Developer Policies and the Ads Developer Blog post of August 31, 2026; Google Ads Help, access levels; Pipeboard's Google Ads MCP guide (updated November 10, 2026) and pricing; Zapier, Google Ads MCP; cohnen/mcp-google-ads; Claude Help Center, custom connectors and connector permissions; OpenAI Help Center, connecting custom MCP servers in ChatGPT; SpaceXSI (formerly xAI), Grok Bot use cases. Google's MCP guide and release notes were rechecked on November 11, 2026; the other pages were read November 11. Google, Pipeboard, Zapier, Anthropic, OpenAI and SpaceXSI (formerly xAI) have no connection to AgentMsgr.

Cost, tokens and the PPC manager