Amazon Ads has an official MCP server, now in open beta — and it's an early signal of where ad tech is heading. Instead of clicking through a console, advertisers point an agent at their account and manage it in plain English. One integration links a custom agent or a mainstream assistant to the ads platform; here's what it does and how to connect it.
What the server exposes
The server turns core Amazon Ads workflows into agent-callable tools. An agent can manage campaigns, check performance reports, adjust account settings, and view billing or financial data — the day-to-day of running paid media, minus the manual navigation. The pitch is speed: ask for last week's spend by campaign, or to pause underperformers, and the agent does the reads and writes for you rather than you exporting spreadsheets.
Connecting it
The headline convenience is that it's one integration covering several clients — you can link a custom-built agent or a platform like Claude, ChatGPT or Gemini to the same server. Because it's hosted by Amazon, you connect it as a remote MCP server and authorise against your Amazon Ads account, so the agent inherits exactly your access. Follow your client's remote-connector flow — the ChatGPT Developer Mode guide and Claude's custom connector both apply — and confirm the current endpoint in Amazon's documentation.
What agentic advertising looks like
This is the same shift MCP brought to engineering, arriving in marketing: the interface stops being a dashboard and becomes a conversation. A marketer asks for a diagnosis — "which campaigns lost efficiency this month, and why?" — and the agent pulls the reports, correlates them, and proposes changes. Amazon shipping an official server (rather than leaving it to third parties) is the tell that vendors now see agent access as a first-class channel, not a side project.
Beta caveats & permissions
It's open beta, so treat capabilities as still moving and expect rough edges. The permission model is the important part: an agent that can adjust campaigns and settings can also spend money, so scope access tightly, keep a human in the loop for budget changes, and review what the agent proposes before it commits. Attribute actions to a service account you can audit.
Going further
Pair it with analytics to close the loop — see best MCP servers for product analytics. For the safety rails, read MCP permission scoping patterns and keeping a human in the loop with elicitation. Marketers assembling a stack should browse the growth-marketer loadout.