Meta Ads MCP in ChatGPT and Gemini: who Meta supports

You pay for ChatGPT or Gemini, not Claude, and a guide tells you Meta’s ads connector works with “any AI assistant”. You paste the address, and either nothing happens or you land on a login screen asking for a client ID you have never seen.

What matters is which path you are on. On Meta’s short list, you sign in with your Facebook login and nothing else. Everywhere else, you bring a Meta developer app of your own. ChatGPT is on the first path. Gemini, today, is on the second.

Meta’s two paths in

From Meta’s documentation, read 14 September 2026, not from connecting an account:

• The supported list: the Business Help Centre article Manage ads from an AI agent with Meta ads AI connectors lists ChatGPT, Claude, Claude Code and Perplexity, each linked to that client’s own setup instructions. It then says that for “an AI agent or platform outside of the four listed above”, you should follow Meta’s developer documentation instead.
• The own-app path: Meta’s developer Get started page (updated 4 September 2026) says owning a Meta app is not required for the listed agents. For everything else, you create an app, add the ads MCP server use case, and pick one of three ways to sign in: OAuth, a user access token, or a system user access token. The token needs at least ads_mcp_management plus either ads_read or ads_management.
• ChatGPT with your own app: the same page tells you to set your app’s redirect URL to match ChatGPT’s specification, choose OAuth, enter the server address, and put your Meta app ID in the OAuth client ID field.

Meta’s server also publishes public sign-in metadata that lists an endpoint where a client can register itself (read 14 September 2026). I did not register a test client, so whether Meta accepts one from an unlisted app is untested. The docs send unlisted agents to the own-app path; plan for that.

The guardrails that apply whatever chat app you use, and the jobs that stay with a person, are what the Claude Ads Operator course is built around.

Where the setup guides go wrong

Client by client

1. ChatGPT on Meta’s list

   OpenAI’s developer mode guide covers the steps. Turn developer mode on in settings, add Meta’s address as an app, sign in with your Facebook login, and then choose developer mode from the plus menu in any conversation where you want to use it. It is web only. OpenAI’s guide warns about “model mistakes on write actions that could destroy data”, and write actions ask for your confirmation by default.

   The trade: ChatGPT decides what counts as a write from a hint each tool carries: “Tools without this hint are treated as write actions.” Meta’s docs do not say which of its tools carry that hint. If the reporting tools lack it, you will be asked to approve plain reads too.

2. Gemini app, through Spark not on Meta’s list

   Google’s custom apps help page says custom connections are “Only available in Gemini Spark”, not in a regular Gemini chat. You also need Spark access, a personal Google Account (not a work or school one), to be an adult in the US, and to have Keep Activity turned on. If a server does not register clients automatically, Google has you enter credentials by hand. For Meta, that means your own app’s ID. Google says Gemini currently asks for manual confirmation on any write.

3. Gemini Enterprise admin sets it up

   Google’s Business edition help page (updated 29 July 2026) says a team administrator connects a custom MCP server. The admin enters an authorization URL, token URL, client ID and client secret, and only the newer streamable HTTP transport works. Meta’s docs connect to its server over HTTP transport, and the client ID would be your own Meta app’s. One detail I could not settle: Meta’s sign-in metadata declares clients with no secret, and I have not tested whether Enterprise accepts one without it.

4. Gemini CLI terminal

   The Gemini CLI MCP docs show gemini mcp add --transport http with an optional Authorization: Bearer header. That header is where an Employee-role system user token would go. The CLI also has a trust setting, which “bypasses all tool call confirmations for this server”.

   The trade: this is the quickest Gemini route to set up, and the only one where a single setting turns off every confirmation.

The chat app decides how you sign in. Meta’s server decides what the agent is allowed to do.

What stays the same in every client

Two protections sit on Meta’s side, so switching chat apps does not change them. Meta’s help article says of ads created through the server: “All ads are paused by default until you set them live.” And it says the rules someone with full control of a business portfolio sets in Business Suite, such as blocking campaign creation or budget edits, are enforced by the server. Where those rules live is covered in the server rules post. Those rules stop an agent even when its chat app never asks you first.

• Reusing your CLI token. Admin-role system user tokens are refused, so you end up debugging a sign-in that was never going to work.
• Turning on trust in the Gemini CLI. Every tool call runs without asking, including ones that create campaigns and change budgets.
• Skipping the Business Suite rules because the client confirms writes. A confirmation you click through protects nothing. A blocked action on Meta’s side cannot be clicked through.

Try this tonight Open Business Suite settings and look under Integrations for Ads MCP server. If it is there, block budget edits on your main ad account before you connect any chat app. If it is not there, Meta’s article says you do not have the feature yet. In that case, keep your first sessions read-only and approve each write yourself. Then add the address to whichever listed client you already pay for.