The model in one paragraph

SpendSignoff is an MCP server. Your AI client — Claude, ChatGPT, or Cursor — talks to it over the Model Context Protocol, and through it reads your Google Ads and Meta accounts and drafts changes. The drafts are proposals: a budget reallocation, a bid change, a pacing fix. None of them touches a live account until you approve it in the SpendSignoff app.

So "AI advertising" here is not an autopilot that spends your money while you sleep. It is a fast reader and a careful drafter. The judgement call — push this live, or not — stays with a human, every time, in V1.

What the AI can and cannot do

The connection requests two OAuth scopes only: mcp.read and mcp.draft. There is no mcp.approve scope, so the boundary is enforced by the protocol, not by a setting you can flip.

• Read spend, campaign structure, budgets, bids, and performance across linked accounts. Reads are ungated on every plan, including Free.
• Draft changes as reviewable before → after diffs via propose_change. A draft is inert — it is a row in a queue, not a pending action.
• Never apply a change on its own. Every live change in V1 is one you approved with the two-step Approve & push live → Confirm control.

Why drafts are the right unit

A chat message is a bad place to keep an ad-account change. It scrolls away, it has no diff, and it has no record. A draft is the opposite: it carries the exact before → after values, the account it targets, a projected impact, and a status you can track. When you approve it, the same object becomes an audit entry.

This is why the AI never edits your account directly even though it could read everything it needs to. The draft is the contract between what the model proposed and what you agreed to spend.

Where to go from here

If you want the precise definitions of a draft, the approval queue, and the autonomy loop, read Concepts. If you want to see the exact tool surface the AI gets, read the MCP tools reference.