Safety
What is an ads MCP server, and is it safe for your ad account?
An ads MCP server lets an AI assistant read and change your ad accounts. Here is how it works, where the risk sits, and four safety models to compare.
· 4 min read
MCP, the Model Context Protocol, is an open standard that lets an AI assistant call tools in other software. An ads MCP server is one of those tool sets for advertising: it sits between your assistant and your ad platform and offers actions such as "list my campaigns", "pull last week's spend by ad group" or "change this daily budget".
That is useful and slightly alarming in the same breath. The assistant stops being a place to ask questions about ads and becomes something that can act on a real account with a real card behind it. This article explains what happens in between, where the risk actually lives, and how different servers handle it.
What an ads MCP server does
Without MCP, an assistant can only work with what you paste into the chat. With MCP it can ask the server for data and send instructions back. The server translates between the assistant and the platform's official API, and it decides which tools exist, what each one is allowed to do and who is allowed to call it.
- Read tools return campaigns, ad groups, ads and performance numbers so the assistant can analyse and report.
- Write tools create or change things: budgets, statuses, ad copy, targeting, whole campaigns.
- Authentication decides whose account the server touches. Usually that is OAuth, where you sign in on the platform and approve a set of permissions, or an API key you create yourself.
Everything interesting about safety follows from those three parts. What can the tools do, who can call them and what checks stand in between?
Where the risk actually sits
The assistant is not the main risk. The risk is a combination of broad access, instant effect and no second pair of eyes.
- Over-broad access. A token that can edit every account you own is far more than a weekly report needs.
- Unreviewed writes. A well-meaning but wrong change, such as a budget typed with an extra zero, goes live the moment it is sent.
- Untrusted text. Assistants read text from many places. Instructions hidden in a web page or a document can try to steer the assistant, which is one more reason a person should see a change before it ships.
- Credentials in the wrong place. Keys pasted into chats or shared documents are hard to take back.
Four safety models you will meet
Servers differ less in what they can do than in how they hold that power back. In practice you will see four approaches, often mixed.
| Model | How it works | Example |
|---|---|---|
| Read-only | No write tools are exposed, so nothing can change. Good for reporting and diagnosis. | Google's open-source Ads server can be run in a strictly read-only configuration, and its first release in October 2025 was read-only. |
| Propose and confirm | A write is described first and only executed after a second, explicit confirmation. | Google's server uses a two-step propose and confirm flow for the few changes it supports, such as assets and labels. |
| Scoped permissions and rules | The platform issues narrow permissions and lets you set rules for what an agent may do. | Meta's hosted ads MCP server uses OAuth with permissions you control, and Meta added rules for agent controls to its Marketing API. |
| Draft first | Every change is stored as a draft, compared with the live account and approved by a person before it is sent. | Adflitz files each change as a draft with a before and after view and guardrail checks. |
None of these is automatically best. A read-only setup is the safest and also the least useful. Draft first is the slowest per change and gives you the clearest record of what happened and why.
Draft first is not the same as paused by default
Two ideas get mixed up. Paused by default is a state on the ad platform: a new campaign is created switched off, so it does not spend until someone turns it on. It limits the damage of a creation that went wrong, but the object already exists in your account.
Draft first works earlier. The change lives inside the tool as a draft. Nothing exists on the ad platform until a person approves it, and the approval screen shows what would change. Adflitz uses both: a new campaign is drafted as paused by default, and the approval step is the safeguard that matters. A draft that asks for something to go live raises a warning for the reviewer.
A safety checklist before you connect anything
- Start with read permission. Ask for reports and audits for a week before you allow a single change.
- Connect one project first, not every account you manage.
- Make sure writes become drafts or confirmations, and that you can see a before and after for each one.
- Look for hard limits that live in software: a cap on budget increases, a maximum number of operations per change, an expiry for unapproved drafts.
- Use one key per tool or script, kept in an environment variable, so a single key can be revoked without breaking the rest.
- Check that there is an activity log that tells you who or what changed what, and when.
- Know where the off switch is, both in the tool and on the ad platform.
If you want to see how this looks in practice, the Adflitz docs on drafts and approvals and access levels show every check in detail.
Sources
Questions people ask
Is an ads MCP server safe?
It is as safe as the permissions you give it and the review step you keep. A read-only connection cannot change anything. A connection that can write is safe when each change waits for a person to approve it, and when budgets and other limits are enforced by software rather than by the AI's good intentions.
Do I need to write code to use one?
No for hosted servers. You paste a URL into an assistant such as ChatGPT or Claude and sign in. Self-hosted open-source servers do need a terminal, credentials and some upkeep.
Can an AI overspend my budget through MCP?
Only if the connection is allowed to change budgets and nothing checks the change. Look for a server that files changes as drafts, shows the before and after, caps budget moves and lets you switch writing off per project.
Can I take access away again?
Yes. OAuth connections can be disconnected and API keys revoked, and you can also remove the permission on the ad platform's own settings page.