Google Ads MCP Server: Set Up Google's Official Server With Claude, and What It Can't Change

Google's official Google Ads MCP server is an open-source Python project at github.com/googleads/google-ads-mcp, published on PyPI as google-ads-mcp. You run it yourself. That can be your laptop or Cloud Run, with Claude Code, Cursor, Gemini or another MCP client pointed at it. It reads your Google Ads reporting data through GAQL queries, and its write tools stop at creating assets, uploading videos and managing labels. It can't create, modify, pause or enable campaigns, and it can't change budgets or bids. For most Shopify brands that makes it a reporting tool with a few low-risk write tools behind a confirmation step. Most of the setup effort is Google Cloud credentials.

Details below come from the GitHub README, the v0.0.6 release of November 2, 2026, and Google's developer integration guide, read on November 16, 2026.

Official server or a community one?

Search "google ads mcp" and you'll find a dozen servers. Google's sits in the googleads GitHub organization under an Apache 2.0 license. The Google Ads API docs link to it, and support questions go to GitHub issues or the #ads-api-ai-tools channel on the Google Advertising Community Discord. The repo was created in October 2025 and reached version 0.0.6 in November 2026, so the version number still signals an early project.

Meta hosts its ads MCP server for you; Google gives you code to run. Community servers each define their own tools, and some may reach further into campaigns and budgets than Google's does, so read the tool list before connecting one to an account that spends money.

What the tools do

The server registers five tool categories. Since v0.0.6, tool names carry their category as a prefix.

Tool (v0.0.6 name) What it does Changes the account?
customers_list_accessible_customers Lists the customer IDs your login can reach No
search_search Runs a GAQL query built from structured fields: customer, resource, fields, conditions, ordering, limit No
metadata_get_resource_metadata Lists the selectable, filterable and sortable fields for a resource such as campaign No
assets_create_asset Creates text, image, YouTube video, sitelink, callout, structured snippet and other assets Yes
assets_upload_video_asset Uploads a local video to YouTube through the Ads API (local mode only) Yes
labels_manage_labels Creates, updates and removes account labels Yes
labels_manage_label_associations Adds or removes labels on campaigns, ad groups, criteria and the account Yes

Google's guide spells out the limit: the server "cannot link assets to campaigns, ad groups, or asset groups", and "cannot create, modify, pause, or enable campaigns, ad groups, ads, budgets, or bids." So an assistant can draft a "Free shipping over $75" callout as an asset, but a person still attaches it to a campaign in Google Ads.

Four read-only resources ship with it too: the API discovery document, metric and segment references, and the latest API release notes. They help the model write valid queries.

Before you install: Google Cloud access

Most of the setup is Google Cloud work. You need a Google Cloud project with the Google Ads API enabled. Enabling the API gives that project Test access, which only works against test accounts. To query your real account, the project needs Explorer, Basic or Standard access; without it, calls fail with CLOUD_PROJECT_NOT_APPROVED_FOR_PRODUCTION.

The old developer token is gone. Google's guide says developer tokens were sunset on September 9, 2026, and the Cloud project that owns your OAuth client now determines the access level. Setup guides written before then still tell you to paste a developer token, and you can skip that step.

You also need Python 3.10 or later, pipx, and the gcloud command-line tool.

Set it up locally with Claude Code or Cursor

  1. Enable the Google Ads API in your Cloud project and create an OAuth client. Download its client secret JSON.
  2. Create Application Default Credentials with the Ads scope:
gcloud auth application-default login \
  --scopes=https://www.googleapis.com/auth/adwords,https://www.googleapis.com/auth/cloud-platform \
  --client-id-file=PATH_TO_CLIENT_SECRET_JSON

The command prints the path of the saved credentials file.

  1. Add the server to your client's MCP config: ~/.claude/settings.json for Claude Code, .cursor/mcp.json for Cursor, .vscode/mcp.json for VS Code. Google's guide pins the version:
{
  "mcpServers": {
    "google-ads-mcp": {
      "command": "pipx",
      "args": ["run", "--spec", "google-ads-mcp==0.0.6", "google-ads-mcp"],
      "env": {
        "GOOGLE_CLOUD_PROJECT": "PROJECT_ID",
        "GOOGLE_APPLICATION_CREDENTIALS": "PATH_TO_CREDENTIALS_JSON",
        "GOOGLE_ADS_MCP_TOOLS_CONFIG": "/absolute/path/to/tools_config.yaml"
      }
    }
  }
}
  1. If you reach the brand's account through an agency or manager account, add GOOGLE_ADS_LOGIN_CUSTOMER_ID with the manager's 10-digit ID, no hyphens.
  2. Restart the client and ask "What customers do I have access to?"

The GOOGLE_ADS_MCP_TOOLS_CONFIG line points to the read-only file in the next section. Leave it out and the server loads its bundled config with every tool enabled.

Google also publishes agent skills for setup and account diagnostics (npx skills add google/skills/skills/ads), which work with Claude Code, Cursor, Codex and Antigravity.

Make it read-only, and the config trap

Every write tool runs a propose-and-confirm loop. The first call is a dry run (validate_only=True) that returns the planned changes and a checksum. The change only commits when the agent calls again with the same parameters, the checksum and validate_only=False, which in practice means after you approve. Google's guide is clear that the checksum "does not replace access control". To block writes outright, turn the write categories off in tools_config.yaml.

When the namespaces block lists anything, every category you leave out is disabled. Google warns that a file containing only labels: false and assets: false turns off the read tools too, leaving you with a server that does nothing. A working read-only file lists all five:

namespaces:
  customers: true
  search: true
  metadata: true
  labels: false
  assets: false

Set its path in GOOGLE_ADS_MCP_TOOLS_CONFIG. The server's working directory under a desktop client may not be the folder you think it is. If that variable points at a missing file, the server refuses to start rather than falling back to defaults, which is the behavior you want.

Using it from claude.ai or ChatGPT

The local setup above works for desktop and command-line clients that start the server as a process. Claude on the web connects to remote servers from Anthropic's cloud, so the server must be reachable over the public internet on port 443. A server on localhost won't work there.

For that you deploy to Cloud Run with an OAuth web client. Google's guide lists the redirect URIs to register: https://claude.ai/api/mcp/auth_callback for Claude and https://chatgpt.com/connector_platform_oauth_redirect for ChatGPT. You also add the adwords scope to the OAuth consent screen and, while the screen is in Testing, add each Google account that will connect as a test user. For production, Google recommends a persistent JWT signing key and a token store such as Firestore or Redis so logins survive restarts. That's a small infrastructure project. If one person on the team asks the questions, the local setup is enough.

Prompts that work for a Shopify brand

  • "List the customers I can access and confirm the account name for 1234567890." Do this first. Most prompts need a customer ID, and naming it in each prompt avoids answers about the wrong account.
  • "Top campaigns by cost over the last 7 days, split by device, with conversions, conversion value and cost per conversion."
  • "For Performance Max campaigns, show conversion value by campaign for the last 30 days against the 30 days before."
  • "Which search terms spent more than $50 in the last 30 days with zero conversions?"
  • "Create a callout asset with the text 'Free returns for 60 days'." The agent will show you a proposal first. Approve it, then attach the asset in Google Ads.

If you want the assistant to see what happened on the site after the click, add Google's GA4 MCP server as a second connection. Our Performance Max guide covers what to look for in those PMax numbers, including brand cannibalization.

Where Google Ads numbers stop

Conversion value in Google Ads is Google's attributed revenue, counted under the conversion actions and attribution model you set up. It won't equal Shopify net sales. Meta, email and organic search claim some of the same orders, and an assistant reading only Google Ads has no way to see that overlap. Shopify attribution explained and choosing an attribution window cover how to compare platforms fairly.

Finsi brings Google Ads, Meta, Shopify, Klaviyo and subscription data into one warehouse with one set of metric definitions, and Finsi MCP gives Claude or ChatGPT read-only, logged access to those numbers. The Google server can't change budgets or bids. Finsi's PULSE agent proposes those changes across Google and Meta with the data behind each one, and a person on your team approves it before anything runs. More on that model in Finsi's AI agents and approval-gated agents for ecommerce.

Sources