Does connecting Meta to Claude let it spend money without my approval?

No. Every change is dry-run first, the MCP shows you a concrete preview of what it will do and waits for your explicit approval before writing anything to your live account. Reads happen freely; writes are always gated.

What Meta role do I need for Claude to make changes?

An admin or advertiser role on the specific ad account. An analyst or read-only role lets Claude report on performance but not edit campaigns, budgets, or creatives.

Is my Meta password shared with the tool?

No. Authorization uses Meta's standard OAuth consent flow. You sign in with Meta directly and grant scoped access; the Ads MCP receives a token, never your password.

Connect Meta Business Manager to Claude

VigilDog Team · October 23, 2026 · 5 min read

Running Meta campaigns usually means clicking through Ads Manager, tab after tab. Connecting Meta Business Manager to Claude changes the shape of that work: you describe what you want in plain language, and an approval-gated assistant drafts the changes against your real account. Nothing goes live until you say so. Here's how the connection works and how to set it up safely.

What connecting Meta to Claude gives you

Once Meta Business Manager is connected, Claude can read and operate your Meta ad accounts through a tool layer rather than the web UI. You can ask it to pull performance, find underspending ad sets, draft a new campaign, adjust budgets, or pause a losing creative, all in conversation, grounded in your actual account data instead of a screenshot you pasted.

The point isn't to hand over the keys. It's to collapse the mechanical parts of campaign management, the reporting pulls, the repetitive edits, the cross-checking, into a single chat while keeping a human decision on every write. This is the same model VigilDog's Ads MCP uses across Google, Meta and LinkedIn.

How an Ads MCP runs a change safely

Before you start: what you need

The connection assumes a few things are already in place. Sorting these out first makes setup a two-minute job instead of a debugging session.

The most common blocker is the account role. Read-only access will let Claude report but not act; if you want it to draft edits, you need an admin or advertiser role on the specific ad account, not just membership in the Business Manager.

A Meta Business Manager with at least one ad account
An admin or advertiser role on the ad account you want to operate
Claude (or ChatGPT) set up to use the VigilDog Ads MCP
A payment method already configured in Meta, the MCP doesn't manage billing

How the connection works

You authorize the connection through Meta's standard OAuth flow, the same consent screen you'd see linking any third-party tool. You sign in as yourself, review the permissions, and grant access to the specific ad accounts you choose. The Ads MCP stores the resulting token securely and uses it to call the Meta Marketing API on your behalf; your Meta password is never seen or stored.

From there, Claude talks to Meta through the MCP's tool layer. When you ask a question, it issues read calls and summarizes what comes back. When you ask for a change, it prepares the write but does not send it yet, which is where the safety model comes in.

The safety model: dry-run and approval gates

Every change is dry-run first. Before anything is written to Meta, the MCP shows you exactly what it intends to do, the campaign it will create, the budget it will set, the ad set it will pause, as a concrete preview. You review that preview and explicitly approve it before it executes. No write reaches your live account on the strength of a chat message alone.

This matters because ad accounts spend real money and a misread instruction can be expensive. The approval gate means an ambiguous request produces a preview you can correct, not a live campaign you have to unwind. If you're new to the pattern, running Google Ads from Claude shows the same dry-run-then-confirm flow on the Google side.

What you can do once connected

In practice the day-to-day looks like a conversation with an analyst who can also execute. A few examples of the kinds of requests that work well:

"Show me every ad set below a 1.5 ROAS in the last 14 days"
"Draft a new prospecting campaign with a $50/day budget and these three creatives"
"Pause the two worst-performing ads in the retargeting set and tell me what you paused"
"Shift 20% of budget from the underperforming ad set to the top one"

If the connection doesn't take

Most connection problems trace back to permissions or scope. If Claude can read but not write, check that your role on the ad account is admin or advertiser, not analyst. If a specific account is missing, confirm it was included when you granted access, OAuth only exposes the accounts you selected on the consent screen, and adding one later means re-running the authorization. And if actions fail with an authorization error after working fine before, the token may have expired or been revoked in Meta's Business Settings; reconnecting refreshes it.

When the connection is healthy, the workflow fades into the background, you stop thinking about which tab a setting lives in and just describe the outcome. VigilDog's Ads MCP brings that same connected, approval-gated workflow to Meta, Google and LinkedIn from one chat.