The Meta Ads MCP hands two Ads Manager jobs to ChatGPT Work: Building a campaign and reading how it’s performing. You describe the campaign you want in the chat and it builds the structure and the ads. Later, when you ask how those ads are doing, it pulls the numbers for you. Here’s how to connect it to ChatGPT Work or Claude, what those two jobs look like in practice, and the risks worth knowing before you point an agent at a live ad account.

What the Meta Ads MCP is

The Meta Ads MCP is a server that Meta hosts. It gives an AI agent access to your ad account through an MCP, the open standard that lets an agent use an outside tool. Behind it sit around 90 tools, grouped into areas like reporting, campaign creation and management, audiences and A/B tests, so most things you’d do in Ads Manager have an equivalent the agent can use.

You won’t find it by searching the plugin directory. You add it as a custom server instead, using a URL that’s the same in every country: https://mcp.facebook.com/ads. Meta explains the setup in its own guide, “Manage ads from an AI agent with Meta Ads AI connectors”. It launches and reads ads, but the images and video have to come from somewhere else, and that’s where Higgsfield comes in.

Add it to ChatGPT Work

It sounds technical, but it only takes a few clicks. In the ChatGPT desktop app, go to Plugins, click Add at the top right and choose Add custom MCP server. Give it a name, paste in https://mcp.facebook.com/ads, leave the authentication set to OAuth and acknowledge the risk. That creates the Meta Ads plugin.

Next, connect it the way you’d connect any account. Click Connect, which sends you to Facebook, and approve the connection in your Ads Manager. Because the connection runs through Facebook Login, the agent can only reach the business portfolios, ad accounts and Pages you grant it, and the rest of your account stays out of bounds.

Claude works the same way: Open Connectors, add a custom connector and paste the URL, then connect through Facebook Login and pick which accounts Claude can reach. If you’d rather follow along with screenshots, a couple of independent setup guides walk through both.

Job one: build a campaign by prompt

Once the connector is live, you can hand it a whole campaign. I asked ChatGPT Work to look up Meta’s best practices on how many ads to start with, recommend the ads it would run, and lay out the structure before it built anything.

The plan it came back with was six ads, each using the 4:5 and 9:16 versions of the creative so every ad displays properly across placements. It also confirmed the account was active and showed me the campaign structure to check. Before it could build, it needed two things from me: The daily budget and the landing page. I gave it 50 euros a day and the URL, and told it to go ahead.

A few minutes later, the campaign, the ad set and all six ads were sitting in my account, with the copy chosen and the images in the right aspect ratios. That’s an afternoon of clicking through menus, done with one prompt. Everything it builds starts paused, too, so nothing goes live until you switch it on. (More on that in the section on risks.)

Job two: pull the performance data

Building campaigns is the obvious use, but reading them back is what turns the connector into a system.

The same connector pulls your ad-level performance from Meta’s Ads Insights API: Spend, impressions and results for each ad, with whatever breakdowns and date ranges you ask for. Once several ads are running in one ad set, Meta’s algorithm moves budget toward the ones that perform. So you’ll see a real spread in how often each ad was shown and how many results it drove, and the agent reads that spread to tell you which creative and which message won.

That closes the loop. The research told you what to make, the build put it live, and now the data tells you what to make more of.

Schedule a weekly analysis

You won’t want to ask for that analysis by hand every week, so set it up once as a scheduled task. Mine runs every Monday at 9am. It pulls the active campaigns, analyses the ad-level data, works out what to test next and posts its findings to our team’s Slack channel.

Scheduled tasks are what turn the connector from a quicker way into Ads Manager into something that keeps improving your campaigns on its own. The weekly analysis is one part of the wider AI ad creative system, and it’s the part that makes everything else compound.

How often you run yours depends on your spend: Daily if you’re running real volume, every couple of weeks if you’re just starting out.

Risks and permissions

Pointing an agent at a live ad account deserves a moment’s care, so it’s worth knowing the guards the connector has built in.

New campaigns, ad sets and ads all start paused, which means the agent can build everything while nothing spends until a person activates it. Access is limited by Facebook Login, so the agent only reaches the accounts you approved. Admins can tighten things further in Business Suite, blocking campaign creation or capping budget changes, so an agent can’t quietly move real money. Meta’s activity log also records every change made through the connector, as independent write-ups confirm. Bear in mind that it’s still in open beta, so some accounts and tools may not be switched on yet.

The sensible approach is the one you’d take with a new team member. Let it do the read-only work freely, and keep a human eye on anything that creates or changes a live campaign until you trust the setup.

Where to start

Connect it, then ask it to pull last month’s performance before you let it build anything, so you get comfortable with the read-only side first. After that, have it draft a small campaign, review the paused ads and activate them yourself.

Everything I use for this, the skills and prompts for the research, the build and the weekly analysis, is free in the 9x community, where our operators share the workflows they actually run. Grab it and point it at one real campaign.

Common questions

What is the Meta Ads MCP? It’s a server hosted by Meta that connects your ad account to an AI agent like ChatGPT Work or Claude, so the agent can build campaigns and pull performance data from the chat. You add it as a custom connector using the URL https://mcp.facebook.com/ads.

Is the Meta Ads MCP a built-in ChatGPT plugin? No, you won’t find it in the plugin directory. You add it as a custom MCP server using Meta’s URL, then connect through Facebook Login and approve which ad accounts the agent can reach.

Can an AI agent spend my ad budget without me? Not by accident. Everything the agent creates starts paused until a person activates it, access is limited to the accounts you approve, and admins can cap or block budget changes in Business Suite. Meta’s activity log records every change made through the connector.

Can the Meta Ads MCP create my ad images? No. The connector launches ads and reads their results, but it doesn’t generate creative. The images and video come from a generation tool like Higgsfield, and the Meta Ads MCP builds the campaign around them.