Claude for Google Ads: What It Can and Cannot Do

Claude can help with Google Ads, but it cannot fully run your account on its own. Through MCP (Model Context Protocol), an open standard that lets AI assistants connect to outside tools and data, Claude can read your campaign data, analyze it and draft changes. Making those changes is a different matter. Google's official server is read-only, and write access comes only from third-party connectors that you have to vet yourself.

This article separates what is documented from what vendors claim, so you can decide how much to hand over.

Can Claude run Google Ads on its own?

No. "Run" suggests an assistant that watches the account, decides and acts. What exists today is closer to a capable analyst sitting next to you.

With a connector in place, Claude can:

• Pull spend, conversions, cost per acquisition and campaign structure into a conversation.
• Compare periods, find wasted spend and explain what changed.
• Audit an account for structure problems, thin ad groups or missing negative keywords.
• Draft headlines, descriptions and keyword ideas for you to review.

Without a write-capable connector, it cannot touch the account. Even with one, the quality of decisions depends on the data it can see. Claude only sees what the Google Ads API or the connector exposes. It is reasonable to assume it has no view into the internal workings of Smart Bidding (Google's automated bid strategies), though you should verify what any given connector exposes.

Be skeptical of claims that an assistant can "fully run" Google Ads. Those come from vendors selling connectors and do not come with methodology.

What does Google's official Google Ads MCP server do?

Google publishes an open-source Google Ads MCP server (repository: github.com/googleads/google-ads-mcp). According to Google's integration guide as quoted in secondary sources, the current release is strictly read-only. It cannot modify bids, pause campaigns or create new assets.

Reported tools:

• list_accessible_customers: lists the accounts your credentials can reach.
• search: runs reporting queries in GAQL (Google Ads Query Language, the SQL-like language the API uses for reports).
• get_resource_metadata: describes the fields available to query.

Older sources list only two tools, so the count has changed over time. Check the repository for the current list.

Three limits matter for merchants:

• It is self-hosted. You run it on your own machine over stdio (a local connection) or deploy it yourself, for example on Google Cloud Run. It is not a Google-hosted endpoint.
• Setup is heavy. You need a developer token from a manager account's API Center, a Google Cloud project and OAuth 2.0 credentials. One source estimates 30 to 45 minutes. MCP does not remove the developer-token requirement or the API quota tied to your access level.
• No announced write support. The "current release" wording suggests writes could come later, but nothing is announced or dated. Do not plan around it.

For the walkthrough, see our guide Connect Claude to Google Ads with the Google Ads MCP Server.

What Claude Google Ads integration options exist?

There is no single built-in integration. Sources disagree on whether Claude's connector directory contains a Google-made Ads app, and none of the research showed a first-party Anthropic one. Your realistic options:

Option	Reads data	Writes changes	Setup	Main trade-off
Google's official MCP server	Yes (GAQL reporting)	No	Developer token, Cloud project, OAuth, terminal	Safe but technical and read-only
Third-party connectors (Supermetrics, Windsor.ai, Markifact, NotFair and others)	Yes, often multi-account	Vendor-stated yes, usually with approval steps	Usually sign-in with OAuth	You trust a vendor with account access; pricing varies
Google's own in-product AI (Ads Advisor, Ask Advisor)	Yes, inside Google Ads	Limited, per Google	None	Not Claude; Google controls what it does

All capabilities in the third row of vendors are vendor-stated. We have not tested them.

Which third-party connectors add write access?

Several commercial products sit between Claude and your ad account. These descriptions come from the vendors' own pages:

• Supermetrics: builds campaigns across Google Ads, Meta, Microsoft Advertising and TikTok from Claude chat. It says every campaign is created paused and every change is logged and reversible. Pricing starts at €39 per month.
• Windsor.ai: says it works on every Claude plan including Free, reads more than 2,500 fields across manager (MCC) accounts, and runs more than 30 write actions with your confirmation.
• Markifact: offers multi-account reporting, purpose-built Google Ads operations, and lower-level GAQL and mutation access.
• NotFair: uses OAuth without requiring you to hold a developer token, exposes 121 tools, and stages bid, budget, negative keyword and ad changes with diffs and undo until you approve. It has a free tier, then a paid plan at $99 a month.

Others named in searches include Ryze AI, Adspirer, Pipeboard, Composio and community servers on GitHub. Pricing and features change, so confirm on each vendor's site.

What is the best MCP for Google Ads?

It depends on what you need. There is no neutral benchmark, so treat any "best" ranking with caution, including vendor blogs.

• You only want analysis and have technical help: the official read-only server. It is free, documented and cannot change anything by design.
• You want to avoid the developer-token process: a hosted third-party connector that handles OAuth for you. Check what the vendor stores and who can see your data.
• You manage many accounts: look for manager-account (MCC) support and per-account permissions.
• You want changes made from chat: pick a connector with staged changes, approval prompts, paused-by-default creation and an undo log. Those patterns appear across several vendors, and they matter more than the tool count.

A caution on similar names: Google also has another repository, google-marketing-solutions/google_ads_mcp, with a different tool set. One guide mentions a flag for enabling mutations, which conflicts with every other source on the official server. Treat that as unreliable.

How can I automate Google Ads with Claude safely?

Start with reading, then add writing in small steps.

1. Begin read-only. Connect the official server or a connector with write tools switched off. Ask Claude for a spend and conversion summary and check it against the Google Ads interface.
2. Limit access. Use a user or account with the narrowest role that works. Prefer a single account over a whole manager account at first.
3. Run an audit. Ask for wasted spend, search terms to exclude and structure problems. Treat the output as a list of hypotheses.
4. Add drafting. Have Claude write ad copy and keyword ideas, and review them yourself against policy and brand rules.
5. Enable writes in a test account. If you use a write connector, try it on a small or paused campaign first. Confirm that changes require approval and that you can see a log.
6. Keep a human sign-off. Approve every change that affects budget or live campaigns, and review the change history in Google Ads afterwards.

For more ideas on tasks that suit this workflow, see Ten MCP Use Cases for E-commerce Marketing Teams.

Where does Claude work with these connectors?

• Claude Code and Claude Desktop: the official server runs locally, and both can launch it from a configuration file. This is the main route for the self-hosted option.
• claude.ai in the browser: one guide says the local server does not work there, because connectors there reach out from Anthropic's cloud and cannot see a server on your laptop. Hosted third-party connectors are designed for this route.

Team and Enterprise plans may have admin controls over which connectors are allowed. Check Anthropic's documentation, which we did not review for this article.

What can go wrong?

• Over-trusting the numbers. A conversational summary can hide a wrong date range, a currency mismatch or a filter. Spot-check against the interface.
• Bad conversion data. An assistant optimizing toward flawed conversion tracking will give confident but poor advice. If your tracking sends revenue rather than profit, or misses orders, fix that first. Some Shopify-focused platforms send profit as the optimization signal, which changes what "good performance" means.
• Vendor risk. A third-party connector holds OAuth access to your ad account. Review its permissions, data handling and how to revoke access. Also check Google's policies on automated tools, which we did not research.
• Unverified performance claims. Figures such as ROAS gains or hours saved on vendor pages have no published method. Do not use them to justify decisions.
• Account-level mistakes. A write tool can change budgets, bids or live ads. Approval gates and undo logs reduce this risk, they do not remove it.
• Overlap with Google's own AI. Google offers Ads Advisor and Ask Advisor, Gemini-based assistants inside Google Ads. Names and beta status have shifted, so check Google's announcements for what is current.

Is this the same as agentic shopping protocols?

No. MCP connects an assistant to tools and data, such as your ad account. Protocols like UCP, ACP, AP2 and A2A concern agents shopping, paying or talking to each other. They do not manage your campaigns.

Bottom line

Claude for Google Ads is a good fit for reporting, audits and drafting, and a riskier fit for making live changes. Start read-only, verify what it tells you, and add write access only with approvals and logs in place. Confirm technical details in Google's developer documentation at developers.google.com/google-ads/api before you build on them.