Meta’s ads connector (mcp.facebook.com/ads, open beta since April 2026) lets an AI assistant such as Claude, ChatGPT or Perplexity read and change the ad accounts your Meta login can reach; what limits it is that login’s access, not what you tell it in the chat, so connect with a login that can only view the accounts it should only read, and have anything it builds created paused and checked by a person.
In April 2026 Meta opened its ads to AI assistants. The Meta Ads AI Connectors, in open beta, include a server at mcp.facebook.com/ads that an assistant can sign in to with your Meta login. Meta names Claude, Claude Code, ChatGPT and Perplexity as supported.
Once connected, the assistant can do far more than report. Meta’s own documentation lists reading performance, activity logs and A/B test results, and also creating and editing campaigns, ad sets and ads, managing catalogues, and setting up tests. In other words: it can spend money.
That is useful, and for an agency it is also the part to set up carefully, because the accounts are your clients’.
Connecting it
- In your assistant, add a connector (some tools call it a custom integration or MCP server) with the address mcp.facebook.com/ads.
- Sign in with Meta when asked, and choose which business portfolio and ad accounts to allow.
- Start a chat and ask for something read-only first, such as last week’s spend and results by campaign, to check it sees the right accounts.
The details differ between assistants and change often while the connector is in beta, so follow the assistant’s own help page for the exact clicks.
The login is the limit, not the chat
This is the part most setup guides skip. An assistant may ask before making a change, and you can tell it to only look. Neither is a lock. What it can actually do is whatever the Meta login it signed in with is allowed to do. If that login can manage campaigns, so can the assistant, including on a misunderstood instruction.
So decide the access before you connect, in Meta’s Business Settings, not in the conversation:
- For reporting and questions, connect with a login that has only View performance on the ad accounts. It can read everything and change nothing.
- If you want it to make changes, use a separate login with Manage campaigns on only the accounts it should touch, rather than your own login with access to every client.
- Keep an account spending limit on each ad account. It is the one cap that holds whatever any person or tool does.
What to let it do, and what not yet
A sensible order is to earn trust in steps:
- First weeks: reading only. Reports, comparisons, “why did this campaign get expensive”. Check its answers against Ads Manager.
- Then: building, but paused. Ask it to create campaigns and ads switched off, and have a person check and switch them on.
- Last, if at all: changing live spend. Budget changes on a running campaign also restart Meta’s learning phase when they are large, so small steps matter. We wrote about how much to raise a budget at a time.
Leave pixels, datasets and Conversions API settings out of it until you are sure. A wrong change there does not show up as a bad day; it quietly corrupts what every campaign learns from.
Ask what the login can do, not what the assistant says it will do.
What it does not do on its own
An assistant connected to Meta answers when someone asks. It does not look at thirty client accounts every few hours unprompted, it does not keep each client’s normal cost in mind between conversations unless you give it that, and when it suggests a change there is no record of who agreed to it beyond Meta’s activity log. For one or two accounts that may not matter. For an agency it is most of the job.