Most Meta ads MCP failures are not the server. They are a sign-in that never finished, a permission set that does not match what you asked the agent to do, or a client that refuses the connector on plan or policy grounds. Work through those three in order before you assume the server is down, because each one produces a different symptom and only one of them is fixed by reconnecting.
Start by separating the two failures
Two symptoms get described the same way and have nothing in common.
The first is a connection that never produced tools. The client shows the server in its settings, no tool list appears, or the chat says it cannot reach it. That is a transport or authentication problem.
The second is a connection with tools that will not do what you asked. The agent can list campaigns but cannot create an ad set, or a write returns an error while reads work. That is a permission or approval problem.
Ask which one you have before changing anything, because reconnect attempts fix the first and do nothing for the second. The Meta Ads MCP guide lists what each of the available servers can read and change, which is the reference you need for the second case.
Cause 1: the OAuth sign-in never completed
Every remote MCP server signs you in the same way: you give the client a URL, a browser window opens, you approve access, and the tools appear in the chat. Meta hosts its ads server at https://mcp.facebook.com/ads and signs in through Facebook Login for Business.
Four things break that flow without an obvious message:
- The URL is wrong for the client. A remote server takes the base URL it documents, not the tools endpoint someone pasted from a forum, and a local server takes a launch command instead of a URL.
- The sign-in window was closed, or the approval never returned to the client. Reconnect and complete it in the same browser profile you use for Meta.
- The client wants a pre-registered OAuth client. Meta’s get started guide expects your own Meta app in that case: Claude Code takes
--client-id <META_APP_ID>and Codex takes--oauth-client-id. Without it, the sign-in starts and never completes. - The Meta account you approved has no role on the ad account you asked about. The connection works, the tools answer, and every query comes back empty.
When you connect through your own developer app, note that Meta added that option on July 16, 2026, together with ads MCP server rules that let whoever controls the business portfolio decide what agents may change. If an agent suddenly loses abilities it had, check those rules before you debug the client.
Cause 2: the permissions do not cover the request
Meta’s get started guide sets the minimum at ads_mcp_management plus either ads_read or ads_management. The difference between the last two is the one that catches people: ads_read lets an app pull ads reporting information, while ads_management lets it read and manage the ad account.
An agent connected with ads_read reports correctly and fails on every create, update or activation. That looks like a broken server and is a scope decision. If you granted only read access on purpose, the fix is to grant the management scope and reconnect, not to retry the prompt.
The same logic applies to custom audiences and other objects. A tool that manages audiences needs the permission that covers them, and a failure there says nothing about the rest of the connection. Why a custom audience is not working covers the account-side causes that survive a correct scope.
Cause 3: the credential expired, or was never stored
Servers that sign you in through OAuth keep working until the grant is revoked or the token expires. Servers that take a raw Meta access token do not.
GoMarble’s server is the clear example in the material: it takes a Meta access token, either passed as a launch argument or created by its installer and saved locally, and when that token expires you generate a new one and update the config. Once it expires, every tool fails at once with an authentication error, and the fix is a new token rather than a reinstall.
Two habits prevent the next outage. Prefer an OAuth server where one exists, because nothing long lived sits in a config file or a URL that ends up in logs. And if you do use a token, keep it revocable on its own, so rotating it does not break anything else in the account.
Codex users can also check the config itself. The server block lives in ~/.codex/config.toml and is shared by the CLI, the IDE extension and the desktop app, so an entry edited in one place applies everywhere.
Cause 4: the client is blocking the connector
A working server can still be invisible because the client will not load it.
In Claude, a remote server is a custom connector, and Anthropic’s help article says they are available on Free, Pro, Max, Ultra, Team and Enterprise plans, with Free limited to one custom connector. If you already use that one slot, a second server cannot be added. On Team and Enterprise an owner adds the connector for the organisation and members connect individually, so an unadded server is an admin gap rather than a fault.
In ChatGPT, a remote server is added from Plugins with Add custom MCP server, where you enter the URL, configure authentication and create it as a plugin. Account and workspace policies apply, so on a business workspace an admin may need to allow it first.
In Codex, the server needs to be registered with codex mcp add and signed in with codex mcp login, or listed in the config file above.
Cause 5: the agent is doing what it was told
The last group is not a failure at all, and it is the most common one in a lead generation account.
Meta’s ad creation tools reference is explicit: write tools create entities in a paused state, and the AI client asks for confirmation before activation. Activation is its own tool, the one that starts spending budget. An agent that built a campaign, three ad sets and six ads has done its job; nothing delivers until someone activates it.
The reference also lists no tool to delete campaigns, ad sets or ads. Only a custom audience can be deleted. An agent asked to clean up an account will report that it cannot, and that is the documented behaviour.
DropCPL’s server follows the same pattern: campaigns, ad sets and ads are created paused, and switching an ad on or off is flagged as a destructive tool. Its consent screen also defaults to read-only, with write access as an explicit opt-in for one client subaccount. So a write that fails on DropCPL usually means the grant is read-only, not that the server is broken.
A checklist to run in order
| Symptom | Check first | Usual fix |
|---|---|---|
| No tools appear | the URL matches the server documentation and the sign-in window completed | reconnect, completing the browser approval |
| Sign-in loops or fails | a client id is required for that server | supply your own Meta app id in the client command |
| Reads work, writes fail | the granted scope is ads_read only | grant ads_management and reconnect |
| Every tool fails after weeks of use | a raw access token expired | generate a new token and update the config |
| The server cannot be added at all | the client plan limit or workspace policy | free a connector slot, or ask an admin to allow it |
| Objects created but nothing delivers | activation is a separate step | activate the entity, or switch the ad on |
| Queries return empty data | the approved account has no role on that ad account | approve access with the right account |
Run this once, top to bottom, and note which row you stopped at. The next time something breaks, the row you wrote down is where you start, which is faster than reconnecting everything and guessing. And keep the account structure simple enough that an agent’s mistakes are cheap: a small lead gen account structure means fewer objects to check when a tool reports something unexpected.