Using Claude and ChatGPT with Meta ads in 2026
Why this topic?
In my previous post, How Meta ads work in 2026, I wrote about how Meta’s own AI decides who sees your ads. While I was researching it, I kept running into a second topic, which is using AI assistants like Claude or ChatGPT to help run the ads themselves.
This became much easier in 2026, because Meta opened an official way to connect these assistants to an ad account. Since then, a lot of videos and articles promise that AI can now do the whole job of a media buyer, which is the person who sets up, watches and adjusts ad campaigns.
I wanted to understand what these tools can really do today, what’s still hype, and where the risks are, so I went through much of what’s been written and recorded about it and collected what still holds true today. It’s written for people who are new to Meta ads, and I’ll keep using it as a checklist for myself too.
The ways to connect AI to your Meta ads
Before going into details, it helps to know one term, which is MCP, short for Model Context Protocol. It’s a standard way for an AI app to connect to another service, a bit like a universal adapter, so that Claude or ChatGPT can read data from that service and do things in it for you. Once a service offers an MCP connection, you can usually add it to your AI app in a few clicks, without writing any code.
From what I read, there are four main ways to bring AI into your Meta ads today, and one way that’s best avoided.
The first is Meta’s own AI assistant, which lives inside Ads Manager and Meta Business Suite. It gives recommendations based on your account and helps with problems like payment errors or a disabled account. In early October 2026, Meta announced that it’s adding the ability to create ads, change targeting and change budgets through a chat, but most of these features were still being tested at that point.
The second is Meta’s official connector for outside AI apps, which Meta calls the Ads MCP server. It went into open beta on April 29, 2026, and it lets you connect Claude, ChatGPT or another AI app that supports MCP directly to your ad account. It’s free on Meta’s side, it needs no coding, and it’s the option most of this post is about.
The third is connectors made by other companies. They do a similar job, but they usually cost money, and they seem to make the most sense if you also advertise on Google, TikTok or other platforms and want to see everything in one place.
The fourth is building your own setup on top of the Marketing API, which is Meta’s official way for software to talk to an ad account. This is the more technical route. People use AI coding tools like Claude Code to write small scripts for them, or automation tools like n8n or Make to build workflows, for example a report that’s created every Monday and sent to the team. Meta also released a command-line tool for developers, which is a program you control by typing commands, together with the connector.
The way that’s best avoided is any tool that logs into Ads Manager and clicks through it for you, like a robot using your mouse. Section 3.2.3 of Meta’s terms doesn’t allow using automated means to access or collect data from its products without Meta’s permission, and breaking Meta’s terms can lead to restrictions on your ad account. The official connector and the Marketing API are the routes Meta allows, so the usual advice is to stick to them and to be patient if access hasn’t arrived yet.
How Meta’s official connector works?
Connecting it takes a few minutes. In Claude, you open the connector settings, choose to add a custom connector, give it a name like Meta Ads and paste Meta’s address, which is mcp.facebook.com/ads. Then you sign in with Facebook, choose which business and ad accounts Claude is allowed to see, and confirm the permissions. A simple first question, like which ad accounts you have access to, shows whether it works. Custom connectors are available on all Claude plans, but the free plan allows only one of them. In ChatGPT, the same address is added as a custom connection in the settings, which is only possible on paid plans, and in company accounts only if the admin allows it. The exact menu names differ between plans.
The connector is still in open beta, and not every ad account has access yet. Meta switches it on account by account, so the sign-in can work while the ad account still shows a message that it isn’t enabled for the Ads MCP. In that case, the common advice is to wait rather than look for a workaround.
Once it’s connected, the AI can do a lot more than read numbers. Meta’s documentation lists about 90 tools in seven groups, which cover performance reports, creating and editing campaigns, ad sets and ads, product catalogs, checks of the Meta Pixel and other tracking, searching Meta’s Help Center, A/B tests, and the history of changes in the account. The AI can also show you a preview of how an ad will look in a specific placement, and in October 2026, Meta announced that the connector can now upload images and videos as well.
There are two layers of control. The first is on Meta’s side, where you choose which ad accounts the AI can use, and where an admin can also set rules in Meta Business Suite about which actions an AI agent is allowed to take. The second is in the AI app itself. Claude asks before using a tool, and you can either approve it each time or allow it always, while ChatGPT asks for confirmation before any action that changes something. The common advice is to always allow the tools that only read data, like reports, and to keep approval on for anything that touches budgets, targeting or ads. If you only want reports, you can turn the editing tools off completely.
Meta says that new campaigns, ad sets and ads created through the connector start paused, so nothing new spends money until you turn it on. Changes to things that are already running are less clear, because Meta’s documentation doesn’t say that they’re held back in the same way. So if you approve a change like doubling the budget of a live ad set, it’s safest to assume it takes effect right away, which is why it’s worth reading exactly what the AI wants to do before you approve anything.
Meta’s connector itself is free, so the only cost is the AI app you use with it.
What people use it for?
From what I read, most of the value right now is in saving time on work that’s slow to do by hand. Here are the five most common uses.
The first is reporting. Instead of exporting data from Ads Manager and building tables, you can ask a question in normal words, like how much each ad set spent last week, what each sale cost, and how that compares to the week before, and get the answer as a summary, a table or a chart. Vague questions are where it goes wrong most often. If it isn’t clear which dates or which numbers you mean, the AI can quietly pick something itself and answer with a figure that looks exact but isn’t, so the usual advice is to be specific and to compare important numbers with Ads Manager.
The second is checking the account for problems. Because the AI can read the whole account at once, it can notice things that are easy to miss by hand, like too many new campaigns that keep restarting the learning phase, ad sets where the same people see an ad again and again, or tracking that isn’t set up properly. For example, it can check whether the Conversions API, which sends sales data from your server to Meta, is switched on, because if it isn’t, Meta misses part of your sales and nobody notices.
The third is regular reports and alerts. Some AI apps can run the same request on a schedule, for example a short summary every morning, or a check several times a day that sends you a message when the cost per result suddenly jumps. In Claude, this works through scheduled tasks in the desktop app, which are available on paid plans and only run while the computer is on and the app is open. You can also ask it to look for your best ads by your own rule, for example ads that get a large share of the budget and stay above your target return for a week or two, so you know which ones to build on.
The fourth is creative work. The AI can write variations of your ad texts and headlines based on the ads that already work, turn a successful video into a script for a new one, or write a brief for a designer or a creator. Claude doesn’t create images itself, so people connect separate image and video tools to it, which I describe in the next section. The important step here is to look at every image before it’s used, because AI can change small details of a product, and an ad that shows a product differently from how it really looks can break Meta’s rules.
The fifth is preparing campaigns. You describe what you want, and the AI sets up a campaign with ad sets and ads, which then waits in Ads Manager, paused, until you’ve checked it. This can save a lot of clicking, but everything still needs a look before it goes live, like the budget, the audience, the placements and the creative enhancements, which are the automatic changes Meta can make to your ads.
There are a few smaller uses too. You can ask the AI to check your ad texts against Meta’s advertising rules before publishing, or to search the Meta Ad Library, which shows the ads other companies are running. Ads that have been running for a long time are often the ones that work, but the library doesn’t show how well they sell, so it’s a source of ideas rather than proof.
All of this works much better when the AI knows your business. On its own, it only sees the numbers in the account, so it doesn’t know your margins, what a customer is worth to you, when you run promotions, or who your best customers are. A week that looks unprofitable can be completely fine, for example, if the goal is to win new customers who’ll buy again later. That’s why the usual advice is to write this context down once, in the project instructions or a saved document, so the AI uses it in every conversation.
Other tools people connect alongside Meta
Meta’s connector only covers Meta, so for many tasks people add a few more tools to the same AI app. Most of them connect the same way as Meta’s connector, which means you either find them in the app’s list of connectors or add them with their address, and then sign in to that service. Most of them also need their own paid plan. These are the ones that seem most useful for Meta ads work.
For images and videos, two common choices are Higgsfield and Arcads. Higgsfield gives the AI access to many different image and video models in one place, and it can also turn a product link into a video ad or make new versions of an existing ad. Arcads focuses on video ads where an AI-made person talks about your product, in the style of videos made by real customers, which is often called UGC. Both work with Claude, and both use the credits from their own subscription. Canva is another option, mainly for simpler graphics, because the AI can create designs, fill in templates and export them as images. It’s available in both Claude and ChatGPT, but it needs a paid Canva plan.
For sales and website data, people connect the places where the sales happen. Shopify has an official connector in Claude’s directory that can read orders, products and sales, so the AI can compare what Meta reports with what the store sold. It can also make changes in the store, so the same care applies as with Meta. Google Analytics doesn’t have such a simple option yet, because Google’s own connector has to be installed on your computer, so a simpler route is a reporting tool like Windsor.ai, which brings Google Analytics, Google Ads and other data into the AI app through one connection.
For sending results somewhere else, the most flexible option is Zapier. Its connector gives the AI access to more than 9,000 apps, so it can, for example, post a weekly report to Slack, add new numbers to a Google Sheet or create a task for a colleague. Each action uses tasks from your Zapier plan.
Every extra tool adds more things the AI can do, but also more places where it can change something, so the same rules apply as with Meta’s connector, which means giving each tool only the access it needs and keeping approval on for anything that creates, edits or sends.
What to keep in your own hands?
The AI is fast, but it doesn’t carry any of the risk, so a person still has to make the decisions that cost money. Here are six things worth watching out for.
The first is mistakes that sound convincing. The AI only knows what’s in the ad account, so when it explains why your results changed, it’s guessing. A competitor who started spending more, a public holiday, or a broken checkout page on your website would all be invisible to it, which is why its explanations are worth checking before you act on them.
The second is the visuals. It isn’t clear yet how much the AI can really see of your images and videos through the connector, because Meta doesn’t explain it in its documentation. Either way, deciding whether a creative is good is still mostly a job for a person.
The third is the learning phase. Every large change to the budget, targeting, bidding or creative can restart it, so an AI that makes small edits every day can keep an ad set from ever settling down. It helps to write this into the AI’s instructions and to make changes in batches rather than one by one.
The fourth is too many ads. Since AI makes new ads cheap to produce, some setups create and launch new versions every week automatically. This can fill an account with weak ads that compete for the same budget, and as I wrote in my previous post, new ads often get almost no budget anyway, so more isn’t automatically better.
The fifth is responsibility. Meta’s new commercial terms, which take effect on October 30, 2026, make the advertiser responsible for what Meta’s own AI features add to an ad or change in it. The same logic applies to anything Claude or ChatGPT writes or publishes for you, because it goes out under your name, so a person should read every text and look at every image before it’s live.
The sixth is data. Connecting an AI app gives it access to your ad account data, so it’s worth checking the app’s privacy settings. If you manage ads for several clients, it’s safer to keep each of them in a separate connection or project, so that one client’s data never ends up in another client’s work.
The safest way to start, from what I read, is to go step by step. First use the connector only for reports, then let it prepare drafts that you check and publish yourself, and only after that consider letting it make changes, with approval turned on.
In short
Connecting Claude or ChatGPT to Meta ads doesn’t replace knowing how Meta ads work, but it takes away a lot of the slow work, like pulling numbers, writing summaries, watching many ad sets at once, and drafting texts and campaigns. What it can’t do is know your business, decide what’s worth spending money on, or take responsibility for what goes live, so those parts stay with you.
These tools are changing month by month, with new features and new connectors appearing all the time, so it’s worth keeping an eye on what Meta and the AI apps announce next.