Google Ads MCP server: Google's official one, setup and limits

A Google Ads MCP server is the piece that lets Claude, or any other MCP client, query your Google Ads account through the API, so you ask questions instead of exporting reports. Google ships an official one: google-ads-mcp, open source under Apache 2.0. It exposes three tools, it is strictly read-only, and you host it yourself with a Google Cloud project and OAuth credentials. If you want to paste a URL into Claude and go, or you want the AI to pause a campaign or add a negative keyword, you need a hosted connector with write access. Here is what the official server does, how to install it, and where each option fits.

For the wider picture across Meta, TikTok, Amazon and Microsoft, see ads MCP servers: which platforms have an official one. This article stays on Google Ads and on what I checked in Google's own documentation on October 28, 2026.

What Google's official server can do

Three tools, plus a few reference resources. The list looks short, but search carries the weight: it runs GAQL, the Google Ads Query Language, so anything the API can read, the model can ask for.

In practice the model turns your question into GAQL, runs it and writes the answer. That gives broad read coverage: search terms, keywords, assets, bidding settings, whatever the API exposes.

What it will not do is stated plainly in Google's docs: the current release is strictly read-only and cannot modify bids, pause campaigns or create new assets. For mutate operations, Google points you to a different tool, the Google Ads API Developer Assistant, which generates client library code. The official path to changing something runs through code, not through the chat.

The September 2026 change: no developer token

Older guides say you need an approved developer token before anything works. That is out of date. Google's MCP page says developer tokens were sunset on September 9, 2026 and are no longer required by the latest version of the server; API access levels now follow the Google Cloud project behind your OAuth credentials.

That removes the slowest step, the approval wait. It does not remove the technical ones.

How to install the official server

You need a Google Cloud project, a terminal, and a Google login that can see the ad accounts.

1. Install pipx. The server is Python and runs with pipx run, with nothing installed permanently.
2. Enable the Google Ads API in your Google Cloud project.
3. Create credentials. The shortest path is Application Default Credentials: create an OAuth client in the project, download its JSON file, then run gcloud auth application-default login with the https://www.googleapis.com/auth/adwords scope and that client file. The command prints the path of the credentials file it saved. Keep it.
4. Configure your MCP client with the pipx command pointing at the repository, GOOGLE_APPLICATION_CREDENTIALS set to that path and GOOGLE_PROJECT_ID set to your project. If you reach client accounts through a manager account, add GOOGLE_ADS_LOGIN_CUSTOMER_ID.
5. Test with a read-only question, such as "list the Google Ads accounts you can see".

Claude Desktop and Claude Code

In Claude Desktop, the server goes into the MCP config file under Settings → Developer. In Claude Code it is one command, in this shape: claude mcp add --transport stdio --env GOOGLE_PROJECT_ID=your-project --env GOOGLE_APPLICATION_CREDENTIALS=path-to-json google-ads -- pipx run --spec git+https://github.com/googleads/google-ads-mcp.git google-ads-mcp. Claude Code is the easiest place to run the official server, because the people who live in a terminal are the ones it was built for.

Claude on the web

The Claude web app cannot launch a local process, so you have to publish the server. The repository documents a Google Cloud Run deployment with OAuth handled by the server, and Google's guide says to add a custom connector in Claude pointing at your Cloud Run URL with /mcp at the end. That is your infrastructure to secure, monitor and pay for. Google also runs a hosted endpoint pilot, open only to allowlisted participants.

Your options side by side

Data pipelines such as Windsor.ai and Supermetrics ship their own MCPs, and GitHub has several community servers. Their read and write scope differs a lot; Windsor.ai MCP server is one worked example.

When the official server is the right call

Pick it when three things are true: someone on the team is comfortable with a terminal and Google Cloud, the job is reading (audits, reporting, account forensics), and Google Ads is the only platform in the question. There it is hard to beat. It is Google's own, it tracks the API, there is no middleman, and GAQL reaches fields no packaged connector exposes.

It stops being enough in three cases:

• The answer needs an action. "Pause that campaign" or "add that search term as a negative" cannot go through it.
• The decision spans platforms. "Where did leads come in cheapest this week, across Google, Meta and TikTok?" needs one connector that sees all three.
• The team is not technical. Asking a media buyer to create an OAuth client and run gcloud is asking them not to use it.

Setting up Vazante as your Google Ads MCP

In Vazante you sign in with Google, tick the Google Ads accounts you want (manager accounts included), then add the connector in Claude under Settings → Connectors → Add custom connector with https://vazante.app/api/mcp. No API key, no Google Cloud project. In Claude Code: claude mcp add --transport http vazante https://vazante.app/api/mcp, then /mcp to authorize. The full walkthrough, with questions to paste, is in connect Google Ads to Claude, and the connector page is Claude connector.

Limits no Google Ads MCP removes

• Bad conversion tracking stays bad. The model reads what the account records. If a tag counts a phone click as a purchase, the answer is wrong and confident.
• Generated queries can pick the wrong field. With the official server the model writes GAQL, and sometimes asks for a metric that does not fit the resource. That is what the metadata tool is for. Check the first answers against the Google Ads interface.
• Your data goes to the model. Google's own README warns that the server exposes your data to the agent or LLM you connect. That holds for every MCP; AI data security for marketers has the questions to ask.
• Google Ad Manager is a different product. If you searched for "Ad Manager" meaning Google Ads, note that Ad Manager is the publisher-side platform with its own API. Neither the official Google Ads server nor Vazante covers it.

Next step

If you are technical and only need reads, install the official server today with the steps above and run a week of audit questions in Claude Code. If you want to ask and act, or see Google next to Meta and TikTok in one answer, try a hosted connector, start read-only in practice, and compare against the Google Ads interface for a few days. New to the protocol? Start with what is MCP for marketers.