MCP for advertising: what it is, and what it changes

An AI assistant is excellent at explaining a table, but it can only work with what it is given. Without a connector, the user exports a CSV, pastes it into the conversation, and the assistant reasons over a frozen snapshot, with no idea which period it covers or what is missing around it.

MCP reverses the mechanism: the assistant asks, the service answers. It can refine, compare two periods, drill down to one campaign — and it always knows where the figure it quotes came from.

The connector gives access to eight read tools. Each returns readings already synced by SaveAds, not a live call to the ad platform — an account connected five minutes ago may have nothing to show yet.

• list_ad_accounts — the connected ad accounts, their platform and their currency.
• get_performance — spend, clicks, conversions and cost per acquisition over a period.
• list_campaigns — campaigns, their budgets and their results.
• list_search_terms — search terms, including those costing money without converting.
• list_recommendations — the recommendations SaveAds produced.
• list_optimizations — what has already been applied, and its measured effect.
• search and fetch — searching within that scope and reading one specific item.

None of the eight tools modifies a campaign, a budget or a bid. An assistant connected to SaveAds can analyse, compare, explain and prepare a decision; it cannot execute one. Changes are applied from SaveAds, where safeguards and rollback exist.

The CRM, included from the Business plan, adds its own tools — and those do write, but inside SaveAds only: log a call, add a note, schedule a follow-up, advance a deal. Every change carries its author (“by the assistant”), can be undone for twenty-four hours, and anything that commits a figure — deal won, amount, merge, deletion — comes back as a proposal you confirm in one click. No campaign, budget or bid is ever touched from the connector, CRM included.

This is not a technical limitation but a deliberate boundary: a model writing directly into an ad account has no log, no undo and no cap.

Every tool call consumes one request from your plan’s monthly AI allowance — the same allowance as the built-in assistant and SaveAds analyses. The response states what is left, so you are warned before the block rather than after.

History depth also depends on the plan: when the returned period is shorter than the one requested, the response says so explicitly instead of implying that data is missing.

Claude and ChatGPT both accept a remote MCP connector, with two different setup paths — ChatGPT requires developer mode to be turned on first, Claude does not. The dedicated pages walk through each one.