Meta Ads MCP servers in 2026: how to talk to your ad account from an AI assistant

This is the most technical subject on this blog, so the first half is written for someone who has never heard the acronym and has no intention of writing code. If you already know what MCP is, skip to the table. If you do not, the short version is that something changed in the last year about how you can interact with your own advertising data, and it is worth ten minutes of your attention even if you decide to do nothing about it.

What is an MCP server, and why would a business owner care?

MCP stands for Model Context Protocol. Strip the name away and it is a plug: a standard shape for connecting an AI assistant to a system that holds real data, so the assistant can look things up and do things there instead of guessing. Before it existed, an assistant knew whatever you pasted into the chat. With it, an assistant can ask your ad account directly.

What that changes in practice is the shape of the question you are allowed to ask. Ads Manager answers questions you can express as a filter and a date range. An assistant with a connection to the same data answers questions phrased the way you actually think them: which of my campaigns got worse after the fifteenth, and was it the same creative each time. That is not a smarter answer than the one in the reports. It is a faster route to it, and the route is most of what stops owners from looking.

What MCP is not: an autopilot, a strategy, or an opinion about your business. It is a connection. Everything you would want around a connection to a live ad account, a record of what things looked like last month, a limit on how much a budget may move in a day, a step where a person sees a change before it lands, sits outside the protocol and has to come from somewhere else.

Does Meta have an official MCP server for ads?

Yes, and it is straightforward to confirm without taking anyone's word for it. On August 22, 2026 a request to mcp.facebook.com/ads answered with a 401 and a correctly formed authorization challenge: a pointer to its own metadata document, and a list of the permissions it wants, which are ads_management, ads_read, catalog_management, business_management, pages_show_list, instagram_basic and ads_mcp_management. That last one is a permission created specifically for this, which tells you the plumbing is deliberate rather than borrowed. The response also carried Meta's own request and trace identifiers, so it is Meta answering rather than something in front of Meta.

The developer documentation is public and loads: an Ads AI Connectors section, an MCP server overview under it, and a separate command line tool. In Meta's own description the server exposes its tools in seven categories, covering reporting, creating and managing ads, catalog work, signals and datasets, help and diagnostics, A/B tests and conversion lift, and activity logs, and it is meant to work with any MCP-compatible agent rather than with one vendor's assistant.

Four things widely repeated about this server could not be confirmed on August 22, 2026, so they are left out rather than passed along: a specific launch date in April 2026, which appears in several blog posts but not in the documentation; a specific count of tools, where Meta's own pages describe seven categories and name no number; that it is free during a beta period; and the claim that anything it creates arrives paused by default. That last one would be an important safety property, and it is exactly the kind of thing to verify yourself on an account you can afford to break before you rely on it.

Which Meta ads MCP server should you use?

There are three real answers plus a category most people overlook, which is that a tool you already pay for may ship one. Repository figures below were read from GitHub on August 22, 2026.

The official server

Start here if you are starting anywhere. It is operated by the company that holds the data, which removes an entire class of question about who else sees your account, and the seven categories of tools cover more than reporting: creating and managing ads, catalog work, datasets, experiments. The thing to read slowly is the permission list. ads_read is the one that looks; ads_management is the one that changes things. A connection that asks for both can create and edit campaigns, and the difference between those two words is easy to scroll past in a consent dialog.

The two open-source servers

Both were created in April 2025 and both are still being worked on, which is not the norm for this kind of project. Pipeboard is the larger of the two at 1,193 stars, was last pushed to on August 19, 2026, and is positioned as a hosted remote server with a free tier and coverage of five ad platforms rather than just Meta. GitHub does not report a recognized license for it, which is worth resolving before you connect it to a company account. GoMarble is smaller at 355 stars, was last pushed to on August 5, 2026, and is MIT licensed and designed to run on your own machine, which is the answer when a client contract or a policy says the data does not leave your infrastructure.

The one you may already be paying for

Worth checking before you build anything, because a line item you already have may cover it. Foreplay includes MCP access on every plan including the $59 Basic tier, which opens your creative library to an assistant. Motion lists full data and MCP access on its $750 Starter plan. Triple Whale ships a read-only server reached with a separate key, and it appears in both the OpenAI and Claude directories. All three connect an assistant to that vendor's data rather than to your ad account, which is a different and usually safer thing.

What does connecting one actually involve?

Less than you would think for the official server, and more than you would like for the self-hosted ones. The shape is the same in every case.

1. 1.Pick the assistant you are connecting. Meta says the server works with any MCP-compatible agent rather than with one company's, so the choice is yours and is mostly about which assistant you already use.
2. 2.Point that assistant at the server. For Meta's own that is a single address. For the open-source ones it is either a hosted address or a process you start on your own machine, which is where a non-developer usually stops.
3. 3.Sign in and approve the permissions, in Meta's own dialog. Read the list rather than clicking through it, because ads_management is write access and it is in the default set.
4. 4.Test on something that cannot break. Ask for last week's spend by campaign, then for frequency by ad set, before you ask for anything that changes state.

What are the risks of handing an assistant your ad account?

Three of them are worth naming plainly, and none of them is a reason not to do it. They are reasons to do it deliberately.

• –Write access looks like read access in a consent screen. The scopes are one word apart. If all you want is to ask questions, there is no reason for the connection to be able to change anything, and it is worth checking whether the client you chose lets you narrow it.
• –An assistant has no memory of your account between conversations unless something outside the protocol gives it one. Ask the same question in two sessions and you can get two answers that were each reasonable on their own, with nothing to reconcile them and no record of which one you acted on.
• –The protocol defines a connection, not a policy. Spending caps, approval steps and change limits are not features of MCP and do not appear because you connected to a well-run server. Whatever guardrails you want around a live budget, you supply them.

Where does MCP stop and a product start?

MCP hands an assistant the keys to an API, and that is a real capability that did not exist for most people a year ago. What it does not hand over is everything around the keys: a record of what the account looked like last month, a rule about how much a budget may move in a day, a step where a person sees a change before it lands, and somebody accountable when it goes wrong. Those are product decisions rather than protocol features, and the protocol was never meant to make them.

Aevin, our own product, sits on the other side of that line, and the honest way to describe the difference is not better or worse but different questions. It connects to Meta through the same kind of permission grant an MCP client asks for, reads the connected accounts every morning, keeps the numbers behind each finding so this month can be compared to last, and shows a proposed change as a before and after that you confirm before anything reaches Meta. Plans are $97, $197 and $497 a month. If you are comfortable at a command line and want direct access to the API through an assistant you already pay for, a raw MCP server is cheaper and far more flexible, and nothing here is trying to talk you out of it. One of the two is access. The other is somebody watching.