Google Ads MCP: How to Connect and Automate Google Ads with AI in 2026
Updated November 2026
The Google Ads MCP is a Model Context Protocol server that lets an AI assistant such as Claude or ChatGPT query your Google Ads account in plain English. You connect it with a developer token and OAuth credentials, and the assistant then pulls reports, runs GAQL queries, and (depending on the server you run) can change campaigns. It works well for analysis. It gets thin when you want an AI to run the account for you.
This guide covers what the server does, how to connect it, what's worth doing with it, where it stops, and when a full Google Ads AI platform is the better choice.
- What the Google Ads MCP is
- Who it's for
- How to connect it
- What you can do with it
- Limits: read vs write scopes and approval flows
- Raw MCP vs Sprites: a side-by-side comparison
- Recommendation vs human-approved execution
- When a full AI marketing agent is the better fit
- Frequently Asked Questions
- Where to start
What the Google Ads MCP is
The Google Ads MCP is a server that exposes Google Ads API functions as tools an AI model can call. MCP, the Model Context Protocol, is an open standard Anthropic introduced in November 2024 so assistants can connect to external data and actions through one consistent interface. The server translates the assistant's request into an API call and returns the result.
Google published its own open-source Google Ads MCP server in October 2025. It launched with two tools: one that lists the customer accounts your credentials can reach, and one that runs GAQL (Google Ads Query Language) searches. That's the reference implementation, and it launched read-only. Community servers and commercial wrappers add mutate operations on top.
The distinction matters because "Google Ads MCP" gets used for three different things: Google's reference server, third-party servers built on the same API, and hosted versions like the one on the Sprites Google Ads MCP page. They share a protocol but not the same capabilities.
Who it's for
Analysts, in-house PPC managers, and agencies with a technical person on staff get the most out of a raw MCP connection. If you already write GAQL, or you're comfortable with a JSON config file and a Google Cloud project, you can be querying accounts within an afternoon.
It's a poor fit for a founder who wants ads running by Friday. The setup assumes you can obtain a Google Ads developer token, create OAuth credentials, and debug a failed authentication. The prompting is on you too. A chat assistant with raw tool access doesn't know your target CPA, your negative keyword history, or which campaigns you've decided to leave alone unless you tell it every session.
Teams that sit in between, say a marketer who runs Google and Meta and wants a similar connection for Meta (the Meta Ads MCP is the equivalent there), usually find that maintaining separate servers per platform becomes its own job.
How to connect it
Every route needs the same three credentials: a Google Ads developer token, OAuth credentials (or application default credentials) with the Google Ads scope, and a Google Cloud project with the Google Ads API enabled. Google reviews developer token applications, so apply before you need it. If you manage multiple accounts through a manager account, you'll also set a login customer ID.
Connect Claude
Claude Desktop runs the official server locally. Add it to your MCP config file:
```json { "mcpServers": { "google-ads-mcp": { "command": "pipx", "args": [ "run", "--spec", "git+https://github.com/googleads/google-ads-mcp.git", "google-ads-mcp" ], "env": { "GOOGLE_APPLICATION_CREDENTIALS": "/path/to/credentials.json", "GOOGLE_PROJECT_ID": "your-gcp-project-id", "GOOGLE_ADS_DEVELOPER_TOKEN": "your-developer-token" } } } } ```
Restart Claude, then ask it to list your accessible accounts. If that returns results, the connection works. Most failures at this stage trace back to an unapproved developer token or credentials missing the Google Ads scope.
Connect ChatGPT
ChatGPT connects to remote MCP servers, not local ones. The official server runs over a local process, so you'd need to host it behind an HTTPS endpoint and register it as a connector (custom connectors require developer mode and are limited by plan). That adds hosting, authentication, and a security review to the job. Hosted servers skip this step, which is the main reason people use them instead of self-hosting.
Connect a custom agent
If you're building your own agent, any MCP client library can attach to the server: the official Python and TypeScript SDKs, the Claude Agent SDK, or the OpenAI Agents SDK. You pass the server as a tool source, and the agent decides when to call it. Custom agents give you the most control, including the ability to add your own guardrails around writes. They also give you the most to maintain. Sprites' MCP connection guide walks through the hosted version if you'd rather skip the infrastructure.
What you can do with it
Reporting is where a raw MCP connection pays off fastest. Ask for spend, conversions, and CPA by campaign for the last 30 days, and the assistant writes the GAQL, runs it, and summarizes the output. It also handles the questions that take three tabs in the Google Ads interface: which search terms spent more than $50 without converting, or which ad groups have a Quality Score below five.
Keyword research is the second use. With search term and keyword data in context, an assistant can cluster queries by intent, spot terms that deserve their own ad group, and draft a negative keyword list. Keep in mind that the server reads your account history. It doesn't give you the forward-looking volume data you'd get from Keyword Planner, so you'd pair it with another source for net-new ideas.
Campaign management is the third, and the one with the most caveats. Servers that include mutate tools can pause campaigns, adjust budgets, add negatives, and change bids. That's powerful and risky in equal parts. A single misread prompt that sets a daily budget to 1000 instead of 100 costs real money before anyone notices.
For teams that want these jobs repeatable instead of prompted from scratch each time, Sprites' Google Ads automation tools package the same reporting and optimization tasks as saved workflows.
Limits: read vs write scopes and approval flows
The first limit is that Google's OAuth scope doesn't separate read from write. The Google Ads API uses a single scope, `https://www.googleapis.com/auth/adwords`, and it covers both. If you assume that "read-only scope" is a setting you can switch on, you'll be surprised.
Real read-only enforcement comes from two places. One is the user's access level inside Google Ads: connect with a user who has Read only access and any mutate call fails at the API. The other is the tool set the server exposes. Google's reference server only offers read tools, so it can't write even with a powerful login. For a first connection, use both: a read-only user and a read-only server.
Approval flows are the second limit, and the protocol won't help you here. MCP defines how a model calls a tool. It doesn't define whether a human signs off first. Claude Desktop prompts you before each tool call by default, which is a rough approval step, but it shows raw parameters, not a plain-language diff of what changes in the account. People click "allow" fast after the tenth prompt.
Three other limits show up in practice:
- Context windows cap how much data one query can carry, so large accounts need aggregated queries.
- API quotas apply, and a chatty agent can exhaust them.
- There's no memory of your account strategy between sessions unless you build it.
Raw MCP vs Sprites: a side-by-side comparison
Raw MCP in a chat assistant is a strong choice for ad hoc analysis, and it's free apart from the model subscription. The comparison shifts once you want work done, not answered.
| Factor | Raw MCP in a chat assistant | Sprites' Google Ads agent |
|---|---|---|
| Setup | Developer token, OAuth, config file or hosted connector | Connect your account and start |
| Interaction | You write prompts and GAQL context each session | Agent works from your goals and account history |
| Scope of work | Queries and, with the right server, individual changes | Research, campaign creation, publishing, and optimization |
| Writes | Depends on server; approval is a click-through on raw parameters | Human-approved execution with plain-language change summaries |
| Platforms | Google Ads only, one server per platform | Google, Meta, LinkedIn, TikTok, Reddit Ads, and SEO |
| Best for | Analysts and technical operators | Marketers who want results without building the system |
Sprites is an AI marketing agent, so it handles the sequence around the API call: finding what to change, building the change, and routing it for your approval. A raw MCP gives your assistant hands. It doesn't give it a plan.
Recommendation vs human-approved execution
Treat AI output on an ad account as two separate things: a recommendation and an executed change. A recommendation is cheap to produce and cheap to ignore. An executed change moves budget. Mixing them is how accounts get damaged.
The safest pattern is that the AI proposes, a human approves, and only then does the system write. That means every proposed change shows up as a readable item: "Pause 4 ad groups with 0 conversions on $812 spend over 45 days," not a JSON payload. You approve, edit, or reject it. Anything you approve gets logged, so you can trace a CPA shift back to the change that caused it.
You can build this on raw MCP, but you have to build it: a queue, a diff view, a permission model, and a rollback path. Sprites ships it as the default, which is why execution always waits for your approval. If you want to see the workflow in practice, the Google Ads optimization use case shows how recommendations become approved changes.
When a full AI marketing agent is the better fit
Pick raw MCP when the job is analysis, you have someone technical, and you're happy to prompt each time. It's flexible and inexpensive, and the official server is a good learning tool.
Pick an agent when any of these describe you:
- You run more than one ad platform and want one workflow instead of one server per platform.
- You need campaigns built and published, not just audited.
- Nobody on the team wants to maintain credentials, hosting, and prompts.
- You want every account write to pass through a human approval step by default.
Most teams I've seen start with raw MCP for a month, find the reporting valuable, and then hit the wall at the point where someone has to turn insights into 40 individual changes. That's the handoff where an agent earns its cost. For teams weighing how far to automate beyond ads, AI marketing automation covers the wider picture, including SEO.
If you'd like a concrete first step, the Google Ads campaign optimizer template runs an audit and returns prioritized changes you can approve.
Frequently Asked Questions
What is the Google Ads MCP?
It's a Model Context Protocol server that gives AI assistants structured access to your Google Ads data through the Google Ads API. Google released an open-source version in October 2025 that supports reading account data with GAQL. Other servers extend it with campaign management tools.
Is the Google Ads MCP free?
The official open-source server costs nothing to run. You still need a Google Ads developer token, a Google Cloud project, and an AI assistant, which may carry its own subscription. Hosting a remote server for ChatGPT adds infrastructure cost.
Can the Google Ads MCP edit campaigns?
It depends on the server. Google's reference server is read-only. Third-party servers can pause campaigns, change budgets, and add negatives, so check which tools a server exposes before you connect a user with edit access.
Is it safe to give an AI write access to my ad account?
It's safe when a human approves each change and the connected user has limited permissions. Start with a Read only user in Google Ads, confirm the output is accurate, and add write access later with spend limits in place. Never connect an unattended agent with full access to a live account.
Does Google Ads MCP work with ChatGPT?
Yes, through a remote MCP connector. ChatGPT can't launch the official local server directly, so you need it hosted at an HTTPS endpoint, or you can use a hosted service. Connector availability depends on your plan.
What's the difference between Google Ads MCP and Google Ads automation tools?
MCP is a connection standard: it lets an AI call functions in your account. Automation tools add the logic around that connection, such as which changes to make, when, and with what approval. An agent like Sprites includes both.
Where to start
If your goal is faster reporting, set up Google's server in Claude this week, connect it with a read-only user, and spend a few days asking it the questions you normally dig for in the interface. You'll know quickly whether it fits how you work.
If your goal is a Google Ads account that gets researched, built, and optimized with you approving each change, connect Google Ads through Sprites and review your first set of recommendations today.