Claude & Meta Ads MCP: what AI agents do in your account

The Meta Ads MCP server is the official connection through which an artificial intelligence agent, such as Claude or ChatGPT, reads and edits a Meta ad account on your behalf. Meta hosts it at mcp.facebook.com/ads and it works with any agent that speaks the MCP protocol. On October 10, 2026 I connected Claude to my ad account through this server and gave it four tasks, without spending a cent. It got as far as a paused campaign, then a safety check stopped it. And in the activity log, everything shows up as done by me.

On October 6, at Advertising Week New York, Meta presented an assistant that will soon create campaigns on request, and it added new features to the server that lets the AI agents you already use work on your campaigns.

I wanted to see what happens when an agent actually gets its hands on an account. Mine, not a client's. The agent can do much of the heavy lifting, but the analysis that comes before any change stays with whoever manages the ads.

What did Meta announce on October 6, 2026?

Meta announced two different things, and it's easy to mix them up.

The first is the Meta AI business assistant, which lives inside Ads Manager. It already exists and gives advice. What's new is a set of features Meta has started testing and that, it says, are coming soon: planning and creating a campaign, generating images and copy, changing audience and budget, scheduling recurring checks. All through a written request, like a chat. According to Meta, advertisers who follow the assistant's suggestions typically get 5% more conversions at the same cost per conversion. It's a figure stated by Meta, without saying over what period and without outside measurement (Meta for Business, October 6, 2026). On the opportunity score recommendations I see things differently, and I get to that further down.

The second is the Meta Ads MCP server, which existed before the announcement: its documentation was updated on July 14, 2026. Here the agent is yours, the one you already use for writing or analyzing data. The server gives it access to reporting, campaigns, catalogs, pixels, A/B tests and the activity log (Meta developer documentation).

October's novelty is that the agent's work lands as a draft in Ads Manager. You and the agent work on the same campaign, and nothing goes live without explicit approval.

How do you connect Claude or ChatGPT to Meta Ads?

You don't need to be a developer. Meta offers two routes: a ready-made connector for AI assistants, which doesn't require creating any app, or your own developer app with an access token (Meta's "Get started" guide). Among the agents with ready-made instructions, Meta lists ChatGPT, Claude, Claude Code and Perplexity, and for all of them the server address is mcp.facebook.com/ads (Meta Business Help Center, "Manage ads from an AI agent").

I used Claude. According to its guide on custom connectors, these are the steps (Claude Help Center):

Open the connectors section (Customize > Connectors) and add a custom connector.
Give it a name and paste the address https://mcp.facebook.com/ads.
Sign in with your Facebook account and authorize the connection.
In each conversation, turn the connector on from the "+" button at the bottom left.

As of October 2026, custom connectors are also available on Claude's free plan, which allows only one. For ChatGPT and Perplexity the address is the same, and the steps are in the guides Meta links from its page.

As for permissions, the minimum is the one for the server itself plus read access to ads; to create campaigns, you need the management permission instead of read access. After that, the agent sees every ad account you have access to, including your clients'. In the connector settings you can also choose whether Claude has to ask you before using a tool. Its guide recommends "Allow always" only for tools you trust to run unsupervised.

The server isn't live everywhere yet, though. On October 10, on the first page of my list, 5 accounts out of 50 replied that the service is being rolled out gradually. One of them is mine. Before designing a workflow around an agent, it's worth checking account by account.

How do you limit what the agent can do?

With the MCP server rules, which Meta puts in the Business Suite settings. If you have full control of the business portfolio, under integrations you'll find "Ads MCP server". There you can block specific actions for each ad account, such as creating campaigns or editing budgets. For budgets there's also a middle ground: blocking only amounts above a cap you choose. If the agent tries, the request is denied (Meta Business Help Center). The same rules work for agencies: a partner with access to the account can set them for the people on its team.

That way the decision stays on your side: the agent reads the numbers and prepares the work, but budgets and new campaigns stay locked until you unlock them.

Then there's the fear of getting the account banned. Meta writes that using the MCP server on its own doesn't put the account at risk: bans come from violations of the advertising standards, such as non-compliant ads or accounts recreated to get around a ban. If it has already happened to you, here's what to do with a disabled ad account.

The test: four tasks, zero euros spent

The agent was Claude, connected through the official connector. I wrote the questions before starting and didn't change them during the test. I had also set two rules: the agent must never use the tool that starts spending, and on existing campaigns it could propose changes but not apply them.

Task	What the agent did	Outcome
Read the account's last 30 days	Read the 25 campaigns with the most spend and Meta's recommendations	Done
Prepare a €20-a-day campaign for the consultation page	Created the campaign, paused	Stopped at the ad set
Say what it would change in current campaigns	Five proposals, none applied	Done
Check the activity log	Found its own change, signed with my name	Done

What does an AI agent see when it reads your numbers?

It sees Ads Manager's numbers, read with a handful of requests. It lined up the 25 campaigns with the most spend in the period. In the last 30 days the account spent €1,065.19 and collected 38 form leads. The average cost per thousand impressions is €4.70. Between the first and third test, four observations came out:

Spend goes almost entirely to the free courses, and the cost per registration varies a lot from one copy to another: €4.51 in the best campaign, €10.50 in the one that spent the most.
In the last 30 days, seven copies of the same courses campaign had spend. Each copy starts over from zero.
The consultation campaign pays €26.95 per thousand impressions, almost six times the account average. It brought no leads from the website.
Two courses campaigns are active at the same time, and the audience of one (Puglia) sits inside the other's (Italy). They risk competing for the same people, even though that day Meta's auction overlap tool flagged nothing, and the Puglia campaign had only just started.

To find these you have to look at many campaigns together, including the ones you haven't opened in weeks, and the agent lines them up in one go.

What does Meta's AI recommend on your account?

Recommendations I find useless, and a score I wouldn't pay much attention to. The same server gives the agent the opportunity score, the number out of 100 that Meta shows at the top of Ads Manager along with the recommendations that, according to Meta, improve performance. Mine is 78 out of 100, with 18 pending recommendations.

Twelve recommendations out of eighteen ask for the same thing: adding AI-generated images, with the promise of a 10% higher click-through rate. Another suggests the pixel's "Purchase" events to spend 14% less per purchase, on an account that doesn't sell products. These are estimates Meta was showing on October 10, calculated "based on our experiment", without saying which one.

Every accepted recommendation raises the score: in the panel on October 11 the first one is worth 3 points, the other 17 together are worth 19. In my view, though, these recommendations are completely useless, and applying them only makes campaign performance worse. Everything goes through analyzing the results, done by whoever manages the account.

The top recommendation can be applied in one click, from the "Apply now" button. An agent connected to your account reads them all, alongside your numbers, and has the tools to edit campaigns. So the instructions you give it have to start from the results. The agent in the test ruled some out on its own, like the "Purchase" events, and for the AI images it asked me (in Europe they also come with transparency obligations).

How far does an agent go if you ask it for a campaign?

In my case, as far as the campaign, created paused and without spending anything. The ad set never existed, because a safety check stopped the agent first.

I wrote to it: "Prepare a campaign to collect consultation requests from davidecosmai.com/consulenza with €20 a day. You decide structure, audience, placements and copy."

After reading the website pixel and the consultation ads already made, it created the campaign with a leads objective and €20 a day, and named it "BOZZA AGENTE AI - CONSULENZA - NON PUBBLICARE" ("AI agent draft - consultation - do not publish"). Meta registered it as paused, as its documentation says: everything an agent creates starts paused, and only a separate tool turns it on (Meta's tool list).

Two details are worth noting. Meta suggested optimizing for the instant form, the one people fill in inside Facebook. The agent ruled it out on its own, because the destination was the website and my Page hadn't accepted the lead form terms. And when it prepared the copy, it left the AI-generated content disclosure to me: Meta's own tool tells the agent never to decide it on the advertiser's behalf.

The next step, the ad set, never happened. Claude Code, the program I use to run Claude, has a safety check in automatic mode that blocks actions on shared resources. The message said "Modify Shared Resources". It didn't ask me for permission: it stopped. The campaign was left without ad sets, so it couldn't spend.

So there were two sets of brakes: Meta's, and those of the program running the agent. Without the second, the agent could have completed the draft with the tools the server gives it.

There's also a limit the agent could have seen on its own, and only looked at afterwards. Between September 13 and October 10, my website pixel recorded the "Lead" event about fifteen times, across all pages. Meta says an ad set becomes "learning limited" when it's unlikely to get about 50 optimization events in the week after its last significant edit (Meta Business Help Center, "About learning limited"). With fifteen in four weeks, the agent's campaign would very likely have ended up there. The agent noticed in the third test, when I asked it what it would change, that is, after it had already created the campaign.

Who shows up in the activity log when an agent works?

You do. In the Ads Manager activity log the row says "Campaign created", €20 daily, changed by Davide Cosmai, October 10 at 11:04 pm. Apart from the campaign name, which the agent chose to mark as a draft, nothing says an agent did it. The only technical trace is in the data the log returns through the API, where the application name shows up: "ads MCP server". In the panel everyone uses, the one with columns, it doesn't appear. If you work alone, it changes little. For an agency or a freelancer managing client accounts it changes a lot: if an agent gets a budget wrong on a client's account, the log says you got it wrong. And if the campaign name doesn't say so, the client has no way of knowing from the history that an artificial intelligence was behind it. Before connecting an agent to client accounts, then, it's worth telling your clients.

What can you hand over to an AI agent in Ads Manager?

The heavy lifting, the work that takes time and prepares decisions without making them. The reporting tools Meta gives the agent read spend, impressions, clicks and conversions at every level of the account, with the breakdowns and date ranges you choose. Others look for unusual patterns or compare your ad sets with those of similar advertisers (reporting tools of Meta's server), and there are also tools for pixel health and A/B tests (server overview and A/B test tools).

Heavy lifting	What the agent does	What stays with you
Reports	Lines up spend and cost per result for every campaign, broken down by age or by placement	Understanding why a number changed, and whether it matters to the client
Checks	Flags odd trends and, as in the test, copies of the same campaign	Deciding whether it's a problem or a deliberate choice
Meta's recommendations	Reads them all next to your numbers	Analyzing the results: in my view, don't apply them
Pixel	Counts which events arrive and how many	Judging whether they're enough for the goal you have in mind
Drafts	Prepares the campaign paused, with structure and copy	Reviewing everything and approving it before spending starts
A/B tests	Prepares the test and reads its results	Choosing what's worth testing

The right-hand column, though, can't be delegated. The agent carries out what you ask it. What to ask, and what to do with what it finds, is decided by whoever knows the account and the business behind it.

What doesn't an AI agent know about your business?

Everything that isn't inside Meta. The agent read the account's numbers, but some things that matter for deciding live elsewhere:

How many course registrations turn into approved course participants. That figure is in my dashboard, not in Ads Manager.
That the consultation page doesn't convert from Google ads either: from October 5 to 10, 12 people arrived from Google Ads and zero requests, according to Google Analytics.
In an online store, which products have margin and which are about to run out.
What the client is asking for right now, for example more leads or better leads.

And even when the data is there, what matters is when you look at it. In the test, the agent created the campaign and only afterwards saw that the pixel was recording about fifteen "Lead" events in four weeks. Someone who does this job runs that check before opening the campaign. It's the analysis that comes before any change, and in the test the agent didn't do it on its own.

Is the Meta AI assistant available in Italy too?

Yes, it's there in my account. It's called "Meta AI business assistant" and opens from the colorful icon at the bottom left of Ads Manager. At the top it carries a "Beta" label.

Under the text field there are four ready-made prompts, and they change from one day to the next. On October 10 one asked for an analysis of a campaign's setup; on the 11th one offered to talk about A/B tests. They're all requests for ideas or explanations, none of them creates anything. I didn't ask it to create a campaign, so I don't know whether it already does that in Italy. Meta, in its announcement, lists those features as coming.

Should you let an AI agent manage your Meta campaigns?

The heavy lifting, yes. The analysis and the decision, no: those go through the experience of whoever manages the ads.

Reading the 25 campaigns with the most spend in the last 30 days at once and finding the useless copies and the overlapping audiences is real help. It's the work that often gets skipped, because it takes time and doesn't feel urgent. An agent does it with a few requests, and the same goes for a weekly report or a pixel check.

The time you save, though, should go into the part that stays with you: understanding why the numbers are what they are and deciding what to change. And for that you also need what happens outside Meta. For example, you need to know whether a lead is worth anything after the form, and that isn't inside Meta. I've also written about why lead generation is about creating opportunities, not handing over closed customers.

In practice, the heavy lifting goes to the agent and the thinking to whoever manages the ads. The tasks split like this:

The agent gathers the numbers and prepares the report, with anomalies highlighted.
You read the report in light of what you know about the business and decide what to change.
The agent prepares the changes you asked for as a draft.
You check the draft and decide whether to publish it.

If you want to try, here's what I'd do:

Connect the agent with read-only permission, at least at first.
Run your tests on an account of your own, with a campaign name that clearly says it's a draft.
After every session, open the activity log and check what changed.
Don't connect it to client accounts without telling them.

Frequently asked questions

What is the Meta Ads MCP server?

It's Meta's official connection that lets an external AI agent, such as Claude or ChatGPT, read and edit a Meta ad account, with the permissions you grant it.

What can an AI agent do on Meta Ads?

It can read reports, look for anomalies, read Meta's recommendations and prepare paused campaigns; analyzing the numbers and deciding what to change stays with whoever manages the ads.

Should you apply the recommendations in Meta's opportunity score?

In my view no: they're completely useless and applying them makes campaign performance worse; give the score little weight, because everything goes through human analysis of the results.

Can connecting an AI agent to Meta Ads get the account banned?

According to Meta, no: using the MCP server on its own doesn't put the account at risk, because bans come from violations of the advertising standards.

Can you stop the agent from touching the budget?

Yes, if you have full control of the business portfolio: in the Meta Business Suite settings, under the ads MCP server, you can block budget changes or only budgets above an amount you choose.

Can an AI agent spend my budget without asking me?

With Meta's server, everything the agent creates starts paused or as a draft. Starting to spend requires a separate activation action, which the agent shouldn't take without your approval.

Do you need a developer app to connect an AI agent to Meta Ads?

No. Meta has a ready-made connector for AI assistants that doesn't require creating an app; the developer app is only for people who want their own connection.

Who shows up as the author of changes made by an AI agent?

The person who connected the agent. In the Ads Manager activity log their name appears, and the application name only appears in the log data read through the API.

Does the Meta AI business assistant create campaigns in Italy?

In my account, in October 2026, the assistant is there in beta and offers ideas and analysis. Meta has announced the features that create campaigns on request as coming.

Read also

More notes on Meta Ads and artificial intelligence:

I gave Claude the keys to my Meta Ads account: what it did on its own