Muze AIShort answer
Yes, if the connector you add exposes more than one account and every call names which account it targets. With Muze you authorize the accounts you want reachable, the assistant lists them before it does anything else, then passes an explicit account ID on each call. The hard part is not capability. It is ambiguity: a vague write instruction applied across several accounts is the easiest way in this category to change the wrong thing.
How it works mechanically
- You authorize accounts individually during the OAuth flow. The connector reaches what you granted and nothing else.
- The assistant calls a tool to list connected accounts before anything else, so it works from real IDs rather than from a name you typed in chat.
- Every call after that carries an explicit account identifier. Meta ad account IDs and Google customer IDs are different shapes, which is a useful accident: it is hard to pass one where the other belongs.
- Adding a platform does not add an account, and adding an account does not add a platform. Those are two separate grants and it is worth keeping them separate in your head.
In Muze the first call is get_my_accounts, and its tool description instructs the model to call it first rather than asking you for IDs. Small design detail, whole class of mistake removed.
What it is genuinely good for
Cross-account work is the strongest argument for connecting an assistant at all, because the alternative is a dozen tabs and a spreadsheet that is stale by Wednesday.
- One question against every account: which are pacing over budget, which have an ad set past target, which have spend and zero conversions this week.
- Finding the account that broke, rather than the account that is loudest in the Monday meeting.
- Running the same review across every account to the same standard, which humans are unusually bad at and machines are unusually good at.
- Consolidated reporting across clients without a separate reporting subscription sitting in the middle of it.
This is the agency case, and it is the one where the time saved is obvious inside the first week.
The failure mode worth taking seriously
A single-account mistake is a bad afternoon. A multi-account mistake is a phone call to a client. The specific danger is a write instruction with a fuzzy scope: pause the underperformers, cut the budget on anything below target, clean up the old campaigns. Every one of those leaves both the threshold and the set undefined, and the assistant will pick reasonable definitions that are not yours.
- Never let a destructive action target a name pattern. "Anything starting with test" will eventually match something that was not a test.
- Name the account in the instruction every time, including when it feels redundant. Especially then.
- Read across accounts freely. Write to one account at a time until the routine is boring.
- Read the account name in the preview first and the change second. The account field is the one people skim.
This is the argument in should I let AI change my ad budget, with the blast radius multiplied by the length of your client list.
An operating pattern that holds up
- One account per conversation for anything that writes. Context bleed between accounts inside a long thread is real and it is hard to spot afterwards.
- A weekly cross-account read as one session, then separate sessions for whichever accounts came back flagged.
- Keep the audit trail per account and review it with fresh eyes, not in the moment you were already inclined to approve it.
- Watch the tool call allowance. Reading a dozen accounts consumes calls quickly, and plans meter calls rather than questions. See pricing.
Setting it up in ChatGPT
- Settings, then Connectors, then Add custom connector. Confirm your ChatGPT plan supports custom connectors before you start.
- Use https://backend.muzecmo.com/mcp as the connector URL.
- Sign in and authorize each account you want reachable, and only those.
- Ask it to list the connected accounts, and check that list against what you expected before you ask it anything else.
Full steps are in the connection guide, and the client-specific notes are on the ChatGPT page.
Related questions
- Can one connection cover several clients?
- Yes. You authorize each account you want reachable and the assistant addresses them by ID. Whether that is appropriate is a client-permission question rather than a technical one, so check what your agreements actually allow.
- Does each account need its own OAuth grant?
- You grant access per platform login, scoped to the accounts that login can reach and that you select. Accounts sitting under a different login need their own grant.
- Does managing more accounts cost more?
- Indirectly. Plans meter tool calls, and reading ten accounts costs roughly ten times reading one. The $99 a month tier at 2,500 calls exists for exactly this case.
- Can it make the same change across all accounts at once?
- Several write calls can be made in sequence, and we would advise against it. Bulk edits across accounts are the least reversible thing you can ask for and the hardest preview to read properly.
- Does this work in Claude as well?
- Yes. MCP is a standard, so the same connector works in Claude, Cursor, Claude Code and other MCP clients without any change.
Keep reading
Other answers
Ask your own account instead
Muze connects Meta Ads, Google Ads, Amazon Ads and Shopify to ChatGPT or Claude, so the answer comes from your numbers rather than a general one. Free to start, and every write previews before it runs.
See how Muze works