Meta Ads MCP
A protocol for handing an assistant real tools instead of screenshots. Useful, and easy to over-permission.
Sevenam installs a Meta advertising system on your own ad account and runs the technology that operates it — written decisions at 7am every morning telling you exactly what to do that day. One fixed setup fee, a fixed monthly, priced to the work.
What it is, in plain terms.
MCP — the Model Context Protocol — is a standard way to give an assistant a set of callable tools rather than a pile of pasted text. A Meta Ads MCP server is one of those tool sets, wrapping the Marketing API so an assistant can ask for campaigns, ad sets, ads, creatives and insights, and get structured data back instead of a screenshot it has to squint at.
The practical difference is that the assistant stops guessing. Ask it why last week was worse and, with a connector, it can pull the actual breakdown by ad set and date rather than reasoning about a number you typed. That is the whole value: it is the difference between an assistant that talks about your account and one that has read it.
Scopes, and the failure modes.
The access token is the whole security story. ads_read answers almost every question worth asking and cannot change anything; ads_management can edit budgets, pause ads and alter targeting. Most people grant the second because a setup guide told them to, then leave it in place. Grant read first, live with it for a fortnight, and add write only when you have a specific job that needs it and a person approving what it does.
The failure modes are dull and predictable. Rate limits, so a chatty agent stops mid-analysis and reports half an answer as if it were whole. Attribution windows, where the assistant compares two figures pulled on different settings and finds a trend that is not there. Currency and timezone, set per ad account rather than per query, which quietly makes a comparison across accounts meaningless. And a token that expires without telling anyone, so the daily read simply stops and nobody notices for a week.
Questions people actually ask.
What is a Meta Ads MCP server?
A tool server that exposes Meta's Marketing API to an AI assistant as callable tools — campaigns, ad sets, ads, creatives and insights — so the assistant can read your account directly rather than working from what you paste into it.
Do I need one to use AI on my ads?
No. You can get a long way pasting exports into an assistant. A connector matters when you want the same read to happen every day without anyone remembering to do it, which is the point at which it stops being a novelty.
What permissions does it need?
ads_read to analyse, ads_management to change anything. Start with read. It answers nearly every question worth asking and it cannot spend your money.
Is it safe?
As safe as the token you give it and the approval step you put behind it. The risk is not the protocol, it is granting write access to something nobody is checking. Scope it narrowly and revoke it from Business Settings when it is idle.
Can it run without me watching?
The reading can, and should — that is the tedious part. Letting it act unattended is a different decision, and we do not recommend it for anything that moves budget.
Do you use one?
Yes. The system reads accounts overnight through a connector and writes the day's decisions with the numbers behind each one. Nothing is actioned until someone agrees to it.
Have your account read the way ours reads them.
Fifteen minutes with Josh, no pitch deck. Bring the account and we will tell you what a daily read would actually surface.
Get started Four fields, twenty seconds →