How do I set up the Google Ads MCP server with Claude?

MCP, the Model Context Protocol, is an open standard that lets an AI assistant call tools another program provides. Google publishes an MCP server for the Google Ads API, open source under the Apache 2.0 license, and you run it yourself: on your computer, or as a web service on Google Cloud Run. It gives the assistant three tools:

• list_accessible_customers returns the Google Ads accounts the Google login you set it up with can open directly.
• get_resource_metadata says which fields a resource, such as campaign, has, and which can be filtered and sorted.
• search runs a query: the assistant names a resource, its fields and its conditions, and the server makes them a Google Ads Query Language query and returns the rows.

It also hands the assistant reference material: the API’s discovery document, the metrics and segments it can report, and the latest release notes. Google says the current release is strictly read-only: it can’t change bids, pause campaigns or create assets.

• A Google Cloud project with the Google Ads API enabled. Enabling it gives the project Test access, which reaches test accounts only. Your real accounts need Explorer access or above, which you apply for on the project’s Google Ads API Overview page, and Google may grant it straight away. Explorer allows 2,880 operations a day against real accounts.
• Credentials for a Google login that can open your accounts. Google’s guide uses Application Default Credentials: an OAuth client you create in the project, signed in through the gcloud command line, or a service account the command line signs in as.
• No developer token. Google retired developer tokens on September 9, 2026. The Cloud project’s access level now decides what it can reach, and the latest version of the server runs without one, so a guide that asks for a token was written before then.
• pipx and the gcloud command line, to run the server and to sign in.
• Your manager account’s customer ID, if you reach your accounts through one. Google calls it the login customer ID, and a call through a manager account without it is refused.

1. 01In the Google Cloud console, choose or create a project, and enable the Google Ads API in it.
2. 02Open the project’s Google Ads API Overview page, expand Upgrade access level, check that the next level is Explorer, and select Apply for access.
3. 03On the Google Auth Platform Clients page, select Create client, choose a desktop app as its type, and download its JSON file. Google shows the client’s secret only once, when it’s made.
4. 04Sign in with the gcloud command line, naming that file, and copy the path it prints after Credentials saved to file.

gcloud auth application-default login \
  --scopes https://www.googleapis.com/auth/adwords,https://www.googleapis.com/auth/cloud-platform \
  --client-id-file=YOUR_CLIENT_JSON_FILE

5. 05Install pipx.
6. 06Add the server to your assistant’s configuration, below, with your project’s ID and the path you copied. If you reach your accounts through a manager account, add GOOGLE_ADS_LOGIN_CUSTOMER_ID to env with its customer ID.

{
  "mcpServers": {
    "google-ads-mcp": {
      "command": "pipx",
      "args": [
        "run",
        "--spec",
        "git+https://github.com/googleads/google-ads-mcp.git",
        "google-ads-mcp"
      ],
      "env": {
        "GOOGLE_PROJECT_ID": "YOUR_PROJECT_ID",
        "GOOGLE_APPLICATION_CREDENTIALS": "/path/to/credentials.json"
      }
    }
  }
}

• Claude’s desktop app: choose Settings from the Claude menu in your computer’s menu bar, then Developer, then Edit Config, and paste the block into claude_desktop_config.json. Quit Claude and open it again to load the server.
• Claude Code: pass the entry inside mcpServers to claude mcp add-json google-ads-mcp, or put the whole block in .mcp.json at your project’s root to share it.
• Cursor: put the block in ~/.cursor/mcp.json for every project, or .cursor/mcp.json in one.

Claude on the web reaches a server from Anthropic’s cloud, and ChatGPT a public address or one through its Secure MCP Tunnel, so neither uses a server started on your computer from a configuration file. Google’s README describes running the server as a web service behind its OAuth proxy, on Cloud Run for example, and connecting Antigravity and Codex to it.

Once it’s connected, Google suggests starting with What customers do I have access to? and How is my campaign performance this week?, and naming the customer ID in a question when you move between accounts.

Google’s server answers what you ask, and nothing more. The model writes every query: search takes a resource, its fields and its conditions, and Google’s own description of the tool tells the model to look fields up rather than guess them. So to find a problem the model has to know it’s there to ask about, know which report shows it and judge what comes back, and the next conversation starts from nothing.

Four of the problems Goldbeater’s checks look for show what that takes:

• Bidding toward a conversion that stopped recording. Each conversion action’s conversions month by month, beside the account’s clicks over the same months, and a judgement of whether a silence is longer than the action’s own rate explains. A 30-day report reads a broken tag, a new action and one that records a conversion a season the same way.
• Paying for searches that never convert. The search terms report over 30 days, each search’s clicks weighed against the rate the campaign’s other searches convert at, so a search with a few clicks isn’t condemned early, and one that converted in that campaign over the year told apart.
• A campaign meeting its target, held back by its budget. Impression share lost to budget, the campaign’s return against its own target CPA or ROAS, tested against chance over the days since its budget last changed, which only change history says, and Google’s budget simulator for what a raise would bring.
• The same keyword in two places. Every enabled keyword’s words and match type across ad groups and campaigns, with broad keywords that differ only in word order, spacing or a plural counted as copies.

Each takes the right queries and a judgement. Written again in each conversation, the window, the join or the threshold can come out differently, or wrong, and the answer won’t say so.

Goldbeater runs 115 checks for 187 problems on every connected account, every 24 hours and whenever you ask. Each finding carries its evidence, the figures and tables behind it, what it’s worth and the fix Goldbeater drafted, and a check such as this one judges by a test on the account’s own data rather than a threshold somebody picked. Here is the first of the four, as the check wrote it on a sample account:

Goldbeater’s own MCP server hands these findings to Claude, ChatGPT or Cursor. You sign in with your Goldbeater account, so there’s no Cloud project, OAuth client or access level to set up, and a call reads what the checks already found rather than querying Google Ads. It only reads, as Google’s does: a change reaches your account when a person approves it in Goldbeater and presses Apply. How to connect Goldbeater’s MCP server.

The two work side by side: Goldbeater’s server for what the checks found, and Google’s for a question they don’t cover.

Is there an official Google Ads MCP server?

Yes. Google publishes one, open source, in the googleads/google-ads-mcp repository on GitHub, and documents it on the Google Ads API developer site. It’s read-only.

Does the Google Ads MCP server need a developer token?

No. Google retired developer tokens on September 9, 2026, and a Google Cloud project’s access level now decides what it can reach. The latest version of the server runs without a token, and Google ignores one sent anyway.

Can the Google Ads MCP server change my campaigns?

No. Google says the current release is strictly read-only: it can’t change bids, pause campaigns or create assets.

Why can’t it read my real accounts?

The Cloud project has Test access, which reaches test accounts only, and Google Ads refuses the call as one from a project not approved for production accounts. Apply for Explorer access on the project’s Google Ads API Overview page.

Can I use it in ChatGPT or on claude.ai?

Once it runs as a web service they can reach, which Google’s README describes, behind its OAuth proxy on Cloud Run for example. Goldbeater’s MCP server runs as one already, at an address you paste into either.