Meta Ads MCP on GitHub: there is no official repo
Every result for this search is somebody else’s project. Meta never published one, and the reason it never will is the same reason its connector needs no token.
Meta has not published a repository for its ads connector, and it is not going to need
one. The official integration is a hosted endpoint you authorise in a browser —
mcp.facebook.com/ads — so there is no source to clone and no token to store. Every
repository that comes back for this search belongs to somebody else. Those are real, several are
good, and all of them ask you for the same thing: a Meta access token you generate yourself and
paste into code Meta did not write.
You searched for a repository because that is how every other developer tool arrives. Find the project, read the README, clone it, run it. It is a reasonable instinct and it points at the wrong shelf here.
I checked before writing this. GitHub's own search, run today: org:facebook mcp in:name
returns a single repository, an eBPF sandbox for constraining agent tool calls, with nothing to do
with advertising. org:facebook ads in:name returns nothing at all. Neither does the
meta organisation.
Why there is no repository to find
Meta's own announcement of the ads connectors explains it in one sentence, and it is the sentence that answers this search:
For MCP, no developer credentials, API setup, or coding required.
That is the whole design. The connector runs on Meta's servers. You add a URL to your AI client, sign in with the Facebook account that already administers the ad account, and Meta's side decides what you are allowed to touch. There is no package because there is nothing to install — the code lives where the data lives. (Source: Meta's ads AI connectors announcement, published 29 April 2026.)
The side effect is the thing that trips people up. Because there is no repository, you cannot read the tool list before you connect. With a self-hosted server you can open the source and see exactly what it will expose. With Meta's, the only way to know is to connect a client and ask it — which is why I had to enumerate it from inside a session rather than from a file. It reported 95 tools when I counted on 18 August 2026, and the same 95 today.
A self-hosted server talks to the Marketing API through a developer app that belongs to you, so its rate limits and access tier belong to you too. Meta's own update to Ads Management Standard Access sets the bar at 500 successful calls in the previous 15 days with an error rate under 15 percent — down from a previous threshold of 1,500 calls. Fall below it and your app drops a tier. Nobody's README mentions this, because on the day you clone it there is nothing to mention.
Whichever route you take, the connection is the easy part — the routine you run on top of it is what changes your week. See the weekly routine →
The three different things GitHub is holding
The results for this search look interchangeable and are not. They are three separate products wearing the same words, and picking the wrong one costs you an afternoon.
-
A server you run yourself MOST OF THE RESULTS
An npm or Python package you launch locally, pointed at the Marketing API with a token you supply. Real software, usually MIT, usually a single maintainer. The right answer if your account is still behind Meta's rollout gate and you need to work today. The trade: you are now operating a service and holding a credential, and both of those are jobs that do not end.
-
A hosted vendor's front door
Some of the highest-ranked repositories are not really the product. The product is a hosted service on the vendor's infrastructure; the repository is the README, the badges and the signup link. Nothing dishonest about that — but you are reading marketing, not the code that will hold your token. Check whether the install instructions end in
npxor in a sign-up page. -
A skill, which is not a server at all
Instructions, not software: a folder of markdown telling the model how to work an ad account through whichever connector you already have. Nothing to run, nothing to authenticate. Mine is one of these — MIT, on GitHub, free — and it is worth being clear that it does not connect anything. If you install it expecting a server you will be disappointed, correctly.
Reading a repository before you clone it
I opened the READMEs of the top results for this search today. The differences that matter are not on the feature table. They are in how the thing wants your token.
| What the README asks for | Where the credential ends up |
|---|---|
--access-token YOUR_TOKEN on the command line | Your shell history, and any process list on the machine |
META_ACCESS_TOKEN in the client config | Plaintext in ~/.mcp.json, which some people sync |
A .env file in the project | Plaintext on disk, one bad git add from public |
| A sign-in page on the vendor's domain | The vendor's database — better hygiene, more trust |
None of those four is disqualifying. The point is that a user token carries whatever permissions your account has, it lasts 60 days unless it is a system-user token that never expires, and the README's install snippet is telling you exactly where it will live. Read that line before the feature list.
The feature count is the least interesting thing in the README. The token line is the most.
- Assuming the top result is the official one. None of them are. Repository names and descriptions are chosen by their authors, and "Meta Ads MCP Server" is not a protected phrase.
- Comparing tool counts across repositories. Every project counts differently — one splits read and write into separate tools, another wraps both. A larger number is a naming decision as often as it is a capability.
- Cloning before checking whether you need to. If Meta's connector is live on your account, the self-hosted route is work you are choosing. There are three situations where choosing it is right, and most people are in none of them.
- Treating an unmaintained repository as inert. It holds a live credential to an account that spends money. An abandoned server is not neutral; it is a dependency nobody is patching.
Try this tonight Before you clone anything, spend two minutes finding out whether you need to. Add Meta's endpoint as a connector in your AI client and ask it to list your ad accounts — the setup is short. If accounts come back, close the GitHub tabs; you are done, and you never generated a token. If you get the not-enabled error instead, now you have a real reason to go shopping, and you know what to look for in the README when you do.
Frequently asked questions
Is Meta's ads MCP server open source?
No. It is a hosted service on Meta's infrastructure, not a distributed package, so there is nothing to inspect or fork. What you can inspect is the tool schema once you are connected — every tool describes itself, and those descriptions are the closest thing to documentation that exists.
Which repository should I use if my account is not enabled yet?
I am not going to rank other people's projects on a page they cannot reply to. Judge them on the same criteria you would any dependency: recent commits, an actual license, more than one contributor, and an install path that does not put a token in your shell history.
Can I self-host Meta's connector on my own infrastructure?
No. The endpoint is Meta's and the authentication goes through Meta. If you have a data-residency requirement that rules out a hosted endpoint, a self-hosted third-party server against the Marketing API is the only route, and you take on the app tier and the credential along with it.
Does a third-party server see more than the official one?
Not inherently. Both go through the same Marketing API, so the ceiling is the same. What differs is which slices each project chose to expose and how carefully — and, importantly, who is enforcing the guardrails: Meta's side enforces some server-side, a self-hosted server enforces whatever its author wrote.
The connector is free. The routine is the part nobody ships.
No repository gives you a way of working. The course is the weekly operator routine on top of whichever connector you end up with — what to read, what to draft, what never to automate.
See what's inside — $67 founding pricePre-sale: the modules are still being recorded and founding members get the checkout link first. No charge today. Public price $97.