Every PPC team is having the same conversation right now. Someone has connected Claude or ChatGPT to an ad account, asked it why CPA jumped last Tuesday, and got a decent answer in ten seconds. Someone else on the team is nervous, because the same connection could, in theory, change a budget. The thing sitting in the middle of that conversation is the Google Ads MCP, and its cousins for Meta, Microsoft and Amazon.
This guide explains what an ads MCP server is in plain terms, what the official Google Ads MCP can and cannot do, how the Meta Ads MCP differs, and how to set one up without handing an AI assistant the keys to your spend. We use a read-only Google Ads MCP connection ourselves while building aubado, so the opinions below come from using it, not from reading the launch posts.
The short version: an ads MCP is excellent for asking questions about your accounts. It is not built for watching them. Knowing that difference is most of what you need.
What is the Google Ads MCP?
The Google Ads MCP is a server that connects an AI assistant to the Google Ads API through the Model Context Protocol, an open standard for giving AI tools access to outside data. Once it is connected, you can ask questions like "which campaigns spent more than 20% over their daily budget this week?" in plain English, and the assistant turns that into an API query, runs it, and answers with your real account data.
MCP itself is not a Google product. It is a shared plug format. The same protocol lets an assistant talk to Google Ads, Meta Ads, Google Analytics, Search Console, Slack or a database. That is why "MCP" suddenly shows up in every ad platform's changelog: each platform publishes one server, and every MCP-compatible assistant (Claude, ChatGPT, Cursor, Gemini-based tools and others) can use it.
The one-sentence definition
An ads MCP server is a translation layer between an AI assistant and an ad platform's API. It does not add new capabilities to the platform. It changes who can reach the data, and how fast.
That last point matters. Everything the Google Ads MCP returns was already available through the API, Google Ads scripts or a well-built report. What changes is that you no longer need to know GAQL, open a report builder or wait for a developer. You ask, and you get an answer.
What the official Google Ads MCP can and cannot do
Google publishes an official, open source Google Ads MCP server on GitHub (googleads/google-ads-mcp) under the Apache 2.0 license. It has been public since October 2025 and is still actively maintained. It exposes three tools:
- arrow_forwardsearch. Runs a Google Ads Query Language (GAQL) query against an account. This is where almost all the value sits: campaigns, ad groups, keywords, search terms, conversions, budgets and so on.
- arrow_forwardget_resource_metadata. Describes which fields exist for a resource such as "campaign", so the assistant can write valid queries.
- arrow_forwardlist_accessible_customers. Lists the account IDs your login can reach.
It also ships reference resources: the API discovery document, the list of available metrics and segments, and the latest API release notes. Those help the assistant write correct queries instead of guessing field names.
It is read-only, and that is a feature
The official Google Ads MCP cannot change bids, pause campaigns, edit budgets or create ads. It reads. Some practitioners see that as its big limitation. We see it as the reason it is safe to use every day. A read-only connection can give you a wrong answer, but it cannot spend your money.
If you want an assistant to make changes in Google Ads, you need a third-party MCP server that wraps the write side of the API. Those exist, some with decent guardrails. Just be clear about what you are choosing: code Google did not write, holding a token that can edit a live account.
| Server | Who builds it | Access | Setup effort |
|---|---|---|---|
| Google Ads MCP (official) | Google Ads developer team | Read-only (3 tools) | Medium: developer token, Cloud project, credentials |
| Meta Ads MCP (official) | Meta | Read and write (29 tools at launch) | Low: add a URL, sign in with Meta Business |
| Google Analytics MCP (official) | Google Analytics team | Read-only | Medium: Cloud project and credentials |
| Community Google Ads servers | Independent developers | Often read and write | Varies; you trust their code with your token |
How to set up the Google Ads MCP
Setting up the official server is not a one-click connector. Plan for 30 to 60 minutes the first time, most of it spent in Google Cloud and the API Center rather than in the AI tool. If you already have a working Google Ads API setup, it is much faster.
Get a developer token with Explorer access
Request it in the API Center of a Google Ads manager account. A token that is only approved for test accounts will return an error on real accounts, so check the access level before you continue.
Enable the Google Ads API in a Google Cloud project
Create or pick a Cloud project and switch on the Google Ads API. Note the project ID; you will need it in the client configuration.
Configure credentials
The simplest route for one person is Application Default Credentials with the Google Ads scope, created through an OAuth client. Teams that want to share one server can run it with an OAuth proxy instead, so every user signs in with their own Google login.
Add the server to your AI client
Claude, Cursor, VS Code and other MCP clients all use the same configuration block: the command that runs the server, plus your project ID, developer token and credentials path as environment variables.
Set a login customer ID if you work through an MCC
If your access runs through a manager account, add the manager's customer ID. Without it, most queries fail with a permission error that looks like a broken setup.
Then test with something boring: "What accounts do I have access to?" followed by "How many enabled campaigns does account 123-456-7890 have?" If both answers match the Google Ads interface, you are ready for real questions. Include the customer ID in your prompts when you work across several accounts. It saves a round of back and forth every time.
Meta Ads MCP: the one that can change things
Meta took the opposite route. Its official Meta Ads MCP launched in open beta on 29 April 2026 as part of Meta's AI Connectors, hosted by Meta itself. You add the connector URL to your AI client and sign in with Meta Business OAuth. No developer token, no Cloud project, no code. It is free during the beta, and Meta has not said what it will cost afterwards.
At launch it exposed 29 tools across four areas: reporting, campaign management, catalog management and signal diagnostics (checking whether pixel and Conversions API events are actually arriving). Campaign management means it can create, pause and edit campaigns, ad sets, budgets and targeting. That is a very different risk profile from Google's read-only server.
Check your Meta permissions today
In August 2026 Meta added granular controls for AI agent access under Business Settings, Integrations. You can limit an agent to specific ad accounts, allow or block individual actions, and block any budget above a threshold you choose. Several sensitive write actions, including budget changes and creating campaigns, are reportedly switched on by default. If anyone on your team connected an assistant since April, open that panel and turn off what you would not let an intern do unsupervised. Meta documents the controls in its help page on managing ads from an AI agent.
Our position is simple: start Meta read-only, even though it can write. Use it for a few weeks to learn where its answers are reliable. Then enable one write action at a time, with a budget cap, and review the activity log Meta now keeps of what an agent actually did. The Ads Manager fundamentals still apply. The assistant does not know your margins, your stock levels or what your sales team promised a client last week.
Other platforms are moving in the same direction. Microsoft Advertising has shipped its own MCP server, and Amazon Ads runs one in open beta with read and write access. Expect LinkedIn and the rest to follow. The pattern is clear: every ad account is becoming something an agent can address.
Five PPC jobs where an ads MCP earns its place
The best use cases share one trait: the hard part is getting the data, not deciding what to do with it. MCP removes the "getting" step. The judgement stays with you.
1. Diagnosing a sudden change
"Conversions on the brand campaign dropped 40% since Thursday. Compare impression share, CPC, conversion rate and search terms for the last 7 days against the 7 days before." One prompt replaces four report tabs. It is the fastest way we know to go from "something is off" to a first hypothesis, and it pairs well with a structured approach to analyzing Google Ads data.
2. Search term mining
Ask for search terms with spend above a threshold and zero conversions over 30 days, grouped by theme, with a draft negative list. You still approve every negative, but the export, filter and pivot step disappears.
3. Questions across accounts and channels
With Google Ads and Meta Ads servers both connected, "compare cost per lead by channel for the last 30 days" becomes one question instead of two exports and a spreadsheet. Check the conversion definitions first. The platforms still count differently, and the assistant will not warn you unless you ask.
4. Meeting prep
Five minutes before a stakeholder call: "Summarise month-to-date spend versus budget per campaign, and flag anything more than 10% off pace." Good enough to talk from. Not a replacement for a proper recurring report.
5. Tracking sanity checks
Meta's signal diagnostics tools and a GAQL query on conversion actions can tell you quickly whether events are still firing. Useful after a site release, a consent banner change or a tag manager publish.
Where the Google Ads MCP stops
Most disappointment with ads MCP servers comes from expecting them to be something they are not. Four limits are worth knowing before you build a habit on top of one.
- infoIt only works when you ask. MCP is request and response. Nothing happens when the chat window is closed. A campaign that blows through its budget at 2 AM will not trigger anything. The protocol has no built-in triggers today.
- infoAnswers are not deterministic. Ask the same question twice and the assistant may write two different queries, with slightly different date ranges or filters. For anything you report upwards, check the query it ran, not just the number it gave.
- infoYour data goes to the model. Google's own README says it plainly: the server exposes your account data to the agent you connect. Check your AI tool's data retention settings and your client contracts before connecting agency accounts.
- infoIt has no memory of your account. The assistant does not know that "brand" means three campaigns, that March always spikes, or that one conversion action is a duplicate. You explain it every session, or you write it into a project instruction file.
The first limit is the important one. MCP is a brilliant tool for asking. It is the wrong tool for watching. Budget pacing, overspend alerts and anomaly checks need something that runs on a schedule, every day, whether or not anyone opens a chat. That is the job of automated rules, scripts, or a dedicated monitoring tool.
It is also the job we built aubado for. aubado checks budgets, pacing and performance across Google Ads, Meta Ads and LinkedIn Ads on its own, and sends you one calm summary each morning. You check once a day, then use an MCP connection when you want to dig into something specific. The two fit together well: one watches, the other answers questions. If that split sounds useful, .
A safe ads MCP setup in seven rules
This is the checklist we would hand any performance marketer before they connect an assistant to a live account. It follows the same logic as any sensible PPC automation strategy: automate the retrieval, keep the decisions.
- checkStart read-only on every platform. Use Google's official server as is. On Meta, block every write action in the AI agent panel for the first few weeks.
- checkPrefer official servers. Community servers are fine for gaps the official ones do not cover, but read the code or at least the issue tracker before handing them a token.
- checkScope accounts tightly. Connect the accounts you actually work on, not the whole MCC or business portfolio.
- checkVerify the first ten answers. Compare them against the platform interface. You will quickly learn which question types it gets right and where it needs more precise prompts.
- checkWrite down your account context. Campaign naming, which conversion actions count, seasonality. Put it in a project instruction so every session starts informed.
- checkEnable writes one action at a time. Pausing an ad is a small risk. Raising a budget is not. Add a budget cap and an approval step before anything that moves money.
- checkReview the change log weekly. Google Ads change history and Meta's agent activity log show what was done and by whom. If you cannot explain a change, turn the write permission back off.
Frequently Asked Questions
Let the MCP answer. Let aubado watch.
aubado keeps an eye on budgets, pacing and performance across your ad platforms every day, and tells you what needs attention before you open a single dashboard. Check once a day. Stay in control.
Related Resources
Google Ads Scripts: What to Automate and What to Skip
The scheduled, unattended side of Google Ads automation, and where it breaks.
StrategyPPC Automation Strategy
Which parts of paid media work to automate first, and which to keep in human hands.
GuideHow to Analyze Google Ads Data
A practical framework for turning account data into decisions, with or without an AI assistant.
Share this article