Google Ads MCP: What It Does and Who It's For

Google Ads MCP is Google's free, open-source server that lets an AI app such as Claude or ChatGPT read your Google Ads data and answer questions about it in plain language. In its current release it is read-only. It can pull reports, but it cannot pause campaigns, change budgets or bids, or create ads.

To run it you need a Google Cloud project with Explorer access or higher, a Google sign-in (OAuth) or a service account, and Python 3.10 or newer (or a cloud host). You no longer need a developer token, because Google retired them on September 9, 2026. It suits someone comfortable in Google Cloud who wants faster answers about their account. It does not suit an owner who wants an AI to run the ads.

Last checked: October 24, 2026. Sources are listed at the end.

What is an MCP server, in plain English?

MCP stands for Model Context Protocol. It is an open-source standard for connecting AI apps to outside systems, which Anthropic open-sourced on November 25, 2024. Many AI apps support it, including Claude, ChatGPT, VS Code and Cursor. The official MCP site compares it to a USB-C port for AI apps. An “MCP server” is the plug on the other end. It tells the AI which actions exist and runs them when the AI asks. Google's server offers three actions: list your accounts, run a report query, and describe a data field. The point to hold on to: the server sets what the AI can do in your account. Each action is a Google Ads API call, so the API's access rules and daily limits below still apply. When the AI calls one of those actions, the results land in the AI's working memory, so the AI model you use sees your account data. For the wider question of letting a chatbot run your ads, see whether ChatGPT or Claude can manage your ad accounts.

Is there an official Google Ads MCP server?

Yes. It lives in Google's googleads/google-ads-mcp repository on GitHub, created on October 3, 2025, under the Apache 2.0 license, so the software costs nothing. On October 28, 2026 it had 1,092 stars. The newest version is 0.0.5, released October 9, 2026. Version 0.0.3 (September 10, 2026) removed the developer token check, and 0.0.5 fixed OAuth access tokens being written to logs. If you set it up months ago, update it.

There are two ways to run it, per Google's developer guide (last updated October 28, 2026):

• On your own computer over “stdio”, where an AI app on the same machine starts the server itself.
• As a web service over Streamable HTTP, for example on Google Cloud Run or any other cloud host. This is the route for Claude.ai and ChatGPT custom connectors. Anthropic says Claude reaches remote MCP servers from its own cloud, so a server that only runs on your laptop cannot be a Claude.ai connector. Claude Desktop can still start a local server from its own config file.

Google's guide also mentions a hosted endpoint pilot for allowlisted participants. The setup page it links to returned a 404 error when we checked on October 23, 24 and 28, 2026, so we cannot tell you who can join.

What the Google Ads MCP server can and can't do

As of October 24, 2026, Google's guide lists the server as “Mode: Read-only (current release)” and says it gives your AI app read-only access to live account data. Google describes its three tools as built for account discovery and performance reporting.

Read-only comes from the tool list, not from the OAuth permission. Google documents one OAuth scope for the Google Ads API, and the server's credentials must carry it. We found no read-only scope, so the permission you grant this server is the same one a server with change tools would use.

Write tools are planned, with no date. On June 4, 2026, a collaborator on the repository wrote of “plans to add write access soon”, each tool switchable in a settings file, tools_config.yaml, that already turns individual tools on or off. The same comment says the project will not accept outside code that adds write features. When write tools arrive, you can leave them off.

Changes still happen by hand. Google's separate Developer Assistant plugin for Claude Code and Google Antigravity writes change scripts but never runs them. Google says to review and run them yourself.

The AI can still get numbers wrong. It writes the report query, and a wrong date range gives a confident wrong answer. Check one number against Google Ads before you trust a report.

What you need to run the Google Ads MCP server

Work down the list. If a line fails, stop there.

1. A Google Cloud project with the Google Ads API turned on. You sign up in Google Cloud Console, no manager account needed. This gives you Test access, which only works on test accounts (practice accounts that never run real ads). Pass: you can create a project in Cloud Console. Fail: you have never opened Cloud Console and do not want to.
2. Explorer, Basic or Standard access on that project. Apply in Cloud Console, not the old API Center page. Google may grant Explorer automatically, with no brand verification. Fail cue: the error CLOUD_PROJECT_NOT_APPROVED_FOR_PRODUCTION means your project is still on Test access.
3. No Free Trial or suspended billing on the project. As of October 24, 2026, Google's developer token page lists two known issues. First, projects on the Google Cloud Free Trial, or with suspended or disabled billing, can be refused Explorer or Basic. Google's workarounds are to upgrade to a paid tier, remove billing from the project, or use another project. Second, some projects that used an old Test-level developer token before September 9 and upgraded after it still get an authorization error on live accounts. Google says a fix is rolling out and suggests a new Cloud project in the meantime. Fail cue: you were upgraded and still get an authorization error on your live account.
4. A way to sign in. An OAuth client ID and secret, application default credentials, or a service account (a login that works without a person signing in each time). If your OAuth consent screen is in Testing, add every Google account that connects as a test user. Google also says an external app left in Testing gets sign-in tokens that expire after 7 days, so expect to sign in again weekly until you publish the app.
5. Python 3.10 or newer and pipx on your computer (the package's PyPI listing requires Python 3.10 or newer, and Google's sample setup starts the server with pipx), or a cloud host such as Cloud Run.
6. An AI app that supports MCP. For Claude.ai or ChatGPT, you host the server and add it as a custom connector with your own OAuth client. Apps without OAuth support get a 401 error and should use the local setup. Anthropic says custom connectors work on Free, Pro, Max, Team and Enterprise plans, with Free users limited to one.
7. A login customer ID, if you use a manager (MCC) account. Set GOOGLE_ADS_LOGIN_CUSTOMER_ID.
8. Not needed: a developer token. Developer tokens were sunset on September 9, 2026. Google says sending one is optional and ignored, and that it will start rejecting developer tokens in a future major version of the API, so remove the token from old setups. Access now comes from the Cloud project that owns your OAuth credentials.

One more rule: Google may revoke API access unused for 90 days in a row.

We have not timed a setup, so we will not give you a number. If every line passes, what is left is a settings file and a sign-in for the local route, or a Cloud Run deployment for Claude.ai and ChatGPT.

Should you set it up? A simple decision rule

• Yes, if you or someone on staff is comfortable in Google Cloud and you want faster answers, such as “which keywords spent money last month with zero conversions?”
• Not yet, if you want the AI to make changes. The official server cannot, and the servers that can carry the risks below.
• Skip it, if line 1 or 2 of the checklist fails and you only look at your account once a month. Copy the numbers you care about into a chat and ask there. No server to set up.

Price is part of the rule. The official server is free. GoMarble's hosted read-only plan, Copilot, starts at $139 a month: $139 × 12 = $1,668 a year for reporting answers. If a free option answers your questions, keep that money in the ad budget.

Worked example: one question, read-only versus write access

This is an illustration, not a test we ran. A plumbing company runs one Search campaign with a $50 average daily budget. The owner asks the AI: “Which keywords spent money in the last 30 days with zero conversions?” With the official server, the AI runs one report query and lists the keywords and their cost. The owner opens Google Ads and removes the junk keywords by hand, because the server cannot.

Now give the AI a server that can change budgets, and suppose it sets $500 instead of $50. Google says a campaign can spend up to 2 times its average daily budget on one day, and up to 30.4 times it in a month:

• Worst normal day: $50 × 2 = $100.
• Worst day after the typo: $500 × 2 = $1,000.
• Monthly ceiling: $50 × 30.4 = $1,520 before, $500 × 30.4 = $15,200 after.
• Extra spend on one bad day versus a typical $50 day: $1,000 - $50 = $950.
• Break-even on gross margin: if the company keeps 40% of each sale after parts and labor, it needs $950 ÷ 0.40 = $2,375 in extra sales to cover that one day. At $400 a job, that is $2,375 ÷ $400 = 5.9, so about 6 extra jobs from a single day's ads.

The read-only limit is a feature for a first setup. A read-only server cannot change your spend. Its worst outcomes are a wrong answer, or your data reaching the AI you connect.

Google Ads MCP servers that can make changes

Several community and paid servers add change tools. The claims below come from each project's own page, checked October 23 and 24, 2026. We have not run any of them.

Two details matter most. Pipeboard's read-only tokens start on its Pro plan, so its $0 tier lacks that lock. And a server's own “read-only” label is only a hint: the MCP specification says those labels are not guaranteed and should be treated as untrusted unless the server is trusted.

What Google's developer rules say about hosted servers

The Google Ads Developer Policies name MCP servers directly. They prohibit a third-party hosted “MCP server, proxy endpoint, or any similar service that solely replicates, wraps, or re-exposes Google Ads programmatic capabilities”. Agencies and advertisers must use “their own Google Cloud Platform account and Google Ads access”. The same page says the rule does not restrict open-source tools that people download and connect “using their own access credentials”, and that end users should sign in by hand to make changes to their accounts. Our reading: running Google's open-source server yourself, with your own Cloud project, is the case that exemption describes. We cannot tell you how Google applies the rule to any named vendor, so before you connect a hosted server, ask whose Google Cloud project carries your requests.

What can go wrong when an AI has write access

1. A budget typo. The worked example above: one wrong zero can raise a day's ceiling from $100 to $1,000.
2. A tool list that is not the real limit. A catch-all tool, like Pipeboard's, can run any Google Ads API change, so the named tools do not show everything the AI can do.
3. A wrong number becomes a wrong change. If the AI misreads a report, a write server lets it act on the mistake.
4. Hidden instructions. Anthropic warns that malicious MCP servers may hide instructions that try to make Claude do things you did not intend, that server makers can change tool behavior without notice, and that it does not verify custom connectors.
5. Nobody watching. A server that changes spend on a schedule removes the person. The MCP specification says a human should always be able to deny a tool call.
6. Your data leaves Google Ads. Google's repository warns that the server exposes your data to whatever AI you connect, read-only or not.

Safety checklist before you connect any AI to Google Ads

1. Start read-only. Use the official server first. When Google adds write tools, keep them off in tools_config.yaml until you need one.
2. Sign the AI in as a limited Google Ads user. Google Ads has five user access levels: Email-only, Billing, Read-only, Standard and Admin. A service account is added as a user at whatever level you choose. Give the AI's login Read-only, never Admin. We have not tested whether the API rejects every change from a Read-only user, so treat it as a second lock, not a guarantee.
3. Set budgets you can live with at 2 times. Google can spend double your average daily budget on a single day.
4. Check change history weekly. Google Ads change history lists account changes from the past 2 years. Changes made through the API can show as “Google Ads API” or the tool's name.
5. Decide who approves. Write down which changes the AI may suggest and which only a named person makes: budgets, pausing, new campaigns. Our guide on which ad tasks to automate and which to keep goes through this.
6. Know how to cut it off. Learn how to remove the user and revoke the app first. Our explainer on what ad account permissions mean covers the steps.
7. Keep the server updated. Old copies carry old bugs, such as the token logging fixed in 0.0.5.
8. Ask any hosted vendor whose Google Cloud project carries your requests. Google ties API access to the Cloud project that owns the OAuth credentials.
9. Try write tools on a test account first. Test access reaches only test accounts, so no real money is at stake.

Google Ads MCP vs Meta Ads MCP

Google's official server only reads. Meta's can change campaigns, so the safety checklist matters more on Meta. Our Meta Ads MCP guide covers that server in detail.

Why other guides still say you need a developer token

Google announced the developer token sunset on September 10, 2026, and pages have not all caught up. On October 24, 2026, the summary box on Google's own quick start still said a developer token is required. The official README still says to apply through the API Center, which Google's sunset page says not to use. The popular cohnen community server and GoMarble's open-source server still walk you through getting a token. One vendor page even quotes wording as Google's that is not on Google's current guide. When pages disagree, trust the body of Google's developer guide and its last-updated date.

If you just want your ads run

An MCP server gives an AI a way into your account. It does not decide what to do with your budget. If what you want is ads written and run for you, that is a different kind of tool. AdFlint is not an MCP server. It runs Google Ads and Meta (Facebook and Instagram) ads inside ad accounts you own. On Google it launches Search campaigns only. You approve every ad before it runs, and budgets have hard caps. Self-serve plans start at $40 a month for up to $1,000 of ad spend, with 0% markup on ad spend. See how it works or read about AI ad management for Google and Meta.

What we could not confirm

• When Google adds write tools. The maintainer said “soon” in June 2026. There is no date and no announced list of what they will change.
• The hosted endpoint pilot. Google mentions it, but the setup page returned a 404 error on October 23 and 24, 2026. We do not know who can join or what it costs.
• What it costs to run. We did not find a Google page that states the API's price plainly, and we did not price Cloud Run hosting. Running it locally needs no hosting.
• Whether a project with no billing account can get Explorer, and how long Explorer approval takes in practice. Google says it upgrades automatically in most cases but gives no time.
• Whether the API blocks every change from a Read-only user. We have not tested it.
• Which ChatGPT plans allow custom MCP connectors. OpenAI's help pages did not load for us, so we have left ChatGPT plan details out.

Questions

Is there an official Google Ads MCP server?

Yes. Google publishes googleads/google-ads-mcp on GitHub under the Apache 2.0 license. It appeared in October 2025, and version 0.0.5 came out on October 9, 2026. You run it on your own computer or cloud host.

Can the Google Ads MCP server change my bids or budgets?

Not the official one. Google lists it as read-only in the current release, with three tools for account discovery and reporting. Some paid and community servers can make changes, including Pipeboard and GoMarble's Operator plan. If you try one, sign it in as a Read-only Google Ads user first.

Do I still need a Google Ads developer token?

No. Google sunset developer tokens on September 9, 2026. Sending one is optional and ignored. Access now depends on the access level of the Google Cloud project that owns your OAuth credentials, and real accounts need Explorer, Basic or Standard.

How do I connect Google Ads to Claude or ChatGPT?

An AI app on your own computer, such as Claude Desktop, can start the official server locally. Claude.ai and ChatGPT reach servers over the internet, so you host the server (for example on Cloud Run) and add it as a custom connector with your own OAuth client. A hosted third-party connector is the other route. Ask whose Google Cloud project it uses.

Why does it say my project is only approved for test accounts?

Your Cloud project is still on Test access, and the API returns CLOUD_PROJECT_NOT_APPROVED_FOR_PRODUCTION. Apply for Explorer in Cloud Console. As of October 24, 2026, Google lists Free Trial billing and suspended billing as known reasons for refusals, plus a bug affecting some projects upgraded after September 9.

Is the Google Ads MCP server free?

The software is free and open source. Other costs depend on your AI app plan and any cloud hosting you choose. On Claude, Free users get one custom connector. Hosted alternatives range from Pipeboard's $0 tier (30 tool runs a week) to GoMarble's Operator plan from $399 a month.

Sources

All checked October 23, 2026, and re-checked October 24, 2026.

• Google, Google Ads MCP server developer guide (last updated October 28, 2026)
• Google, googleads/google-ads-mcp repository, releases and issue #84 (June 4, 2026)
• google-ads-mcp on PyPI (version 0.0.5, requires Python 3.10 or newer)
• Google, API access levels (last updated October 28, 2026)
• Google, developer token sunset and known issues (last updated October 28, 2026)
• Google, brand verification (last updated October 28, 2026)
• Google Ads Developer Blog, new onboarding experience for the Google Ads API (September 10, 2026)
• Google Ads Help, Google Ads Developer Policies
• Google, using OAuth 2.0 to access Google APIs (7-day refresh tokens for apps in Testing)
• Google, OAuth internals and service accounts (last updated October 28, 2026)
• Google, Developer Assistant (last updated October 28, 2026)
• Google Ads Help, access levels in your Google Ads account, change history and average daily budgets
• Model Context Protocol, introduction and tools specification and schema (2025-06-18 version)
• Anthropic, introducing the Model Context Protocol (November 25, 2024) and custom connectors using remote MCP
• Meta, ads MCP server overview, get started, rules and ads AI connectors announcement (April 29, 2026, updated for July 16, 2026)
• Vendor and community pages: GoMarble MCP, GoMarble open-source server, Pipeboard write access guide, Pipeboard pricing, Coupler.io, Adalysis, cohnen/mcp-google-ads and the Glama directory