ESEN
Blog
ESEN
Blog
← All articles

What Is an Amazon MCP? What Amazon Offers Today for Seller Central and Ads

What MCP is, what an MCP server linked to your Amazon account does, what Amazon offers today for Ads and Seller Central, and how access is granted and revoked.

By AGTA Team

An Amazon MCP is a server that uses the Model Context Protocol so an AI assistant such as Claude or ChatGPT can query your seller or advertising account and, if you allow it, make changes there. Amazon currently has two official pieces: the Amazon Ads MCP Server, in open beta since February 2, 2026, and the Selling Partner plugin, which connects Amazon Quick and Claude to Seller Central. There are also third-party servers built on the same APIs.

What MCP is

MCP (Model Context Protocol) is an open standard for connecting AI applications to external systems. Its documentation compares it to a USB-C port: one standard way to plug data sources and tools into the AI app you already use (modelcontextprotocol.io).

There are two sides. The client is the AI application. The server offers tools, each with a name, a description and the inputs it needs. The model reads that list and picks which tool to call based on what you asked. The specification says tools are designed to be invoked by the model on its own, which is why it recommends always keeping a human who can deny a tool call, and having the app ask for confirmation before running operations (MCP specification, tools).

What an MCP server for an Amazon account is

It's an MCP server whose tools work against your account. Under the hood it talks to Amazon's interfaces: the Selling Partner API (SP-API) for Seller Central and the Amazon Ads API for advertising. The assistant doesn't log into Seller Central the way you do. It calls tools like "get this week's sales" or "change this campaign's budget," and the server turns that call into the matching API request.

That leads to the two questions worth asking before you connect anything: which tools the server exposes (read only, or write too), and what permission was used to authorize access to your account.

What Amazon offers today

The Amazon Ads MCP Server

Amazon announced the open beta on February 2, 2026. According to the announcement, it's available globally to Amazon Ads partners with active API credentials, and it turns natural-language instructions into Ads API calls. It can create, update and delete campaigns, run performance reports, manage account settings and pull billing data (Amazon Ads).

The technical documentation lives in the API portal (Amazon Ads API, Overview). The parts a seller should know:

  • There are standard tools, which carry out specific actions (for example, updating a budget or building a full campaign in several steps), and "expert" tools, which interpret the request. The first expert tool writes SQL queries for Amazon Marketing Cloud and is in closed beta.
  • Amazon builds and maintains the server but doesn't operate the models that connect to it. It lists Claude, ChatGPT, Amazon Kiro, Amazon Bedrock and Amazon AgentCore as possible clients.
  • To connect you need a Login with Amazon application with Ads API access and an access token, or OAuth 2.1 authorization through the browser. There's one endpoint per region: North America, Europe and Far East (Amazon Ads API, connecting to the server).
  • You can turn tools off in the client configuration. Amazon's own example disables the delete-campaign and delete-ad tools.

Our read is that, for now, the people who get the most out of it already work with the Ads API: agencies, software companies, or advertisers with someone technical on the team.

The Selling Partner plugin for Seller Central

This is the piece aimed at sellers. Its help page says it connects your Amazon business data to an AI assistant and works with Amazon Quick and Claude (Seller Central, Selling partner plugin for AI assistants). The page applies to selling in the United States, and the plugin is in beta with a gradual rollout, so you may land on a waitlist when you connect.

It works with sales and revenue, inventory, traffic and conversion, listing health and optimization opportunities. Amazon's examples include finding ASINs at risk of stocking out, finding suppressed listings, estimating fees at a given price, starting an FBA inbound plan, and updating the bullet points on a SKU. For that last one, the assistant proposes the change and waits for your approval.

To authorize it you sign in with a seller account on the Professional plan. The account owner or an administrator decides on the Manage Agents page which secondary users can authorize it.

Amazon Ads MCP ServerSelling Partner plugin
Datacampaigns, reports, accounts, billingsales, inventory, traffic, listings, fees, inbound shipments
Statusopen beta since February 2, 2026beta with a waitlist
Who uses itpartners with active Ads API credentialssellers on the Professional plan (US help page)
Clients Amazon namesClaude, ChatGPT, Kiro, Bedrock, AgentCore and other compatible appsAmazon Quick and Claude
Writescreates, updates and deletes campaignsproposes changes you approve; can be limited to answers

Our news piece on the holiday season covers how Amazon suggests using these tools before, during and after peak.

How access is granted and revoked

In Seller Central (SP-API)

Public applications are authorized with OAuth 2.0 through Login with Amazon, and the seller is the one who authorizes. Authorization is renewed every year, or sooner if the app asks for new roles (SP-API, Authorize Applications). To revoke it: Seller Central, Apps and Services, Manage Your Apps, find the app and choose Disable authorization. Amazon notes that only the selling partner can revoke an OAuth authorization (SP-API, Revoke Authorizations).

In the Selling Partner plugin

You manage it in Seller Central under Permissions, Manage Agents, next to Manage Your Apps. You can revoke access there at any time, but Amazon warns that revoking doesn't reverse earlier actions or data access that already happened.

In Amazon Ads

In your advertising account, under Access and Settings, the Third Party Applications tab shows which apps have access and who authorized them. An app gets the same access level as the user who authorized it: a user with view-only access can only delegate view-only access. Once you remove an app it loses access, but it may keep data it already collected, and Amazon suggests asking the provider to delete it (Amazon Ads, third-party applications).

What to keep an eye on

Tools that write

Changing a bid, pausing a campaign or editing a listing takes effect in your account right away. The Ads server can even delete campaigns, and the client lets you switch those tools off. Before you connect, look at the tool list and disable what you won't use.

Approval for each change

The plugin asks for your approval by default and can be limited to answering questions. The MCP specification recommends the same for any client: confirmation before sensitive operations, and showing you the inputs a tool is about to send.

Where your data goes

The plugin page says that once you authorize it, your account data is processed outside Amazon under your AI assistant's terms, and that you're responsible for every action the agent takes, whether or not you started it. It recommends opting out of model training, granting only the minimum permissions you need, and disconnecting agents you no longer use.

How fresh the data is

If a number from the assistant doesn't match Seller Central, Amazon suggests checking that you're comparing the same date range and the same store, and asking again to refresh the data. Always ask the assistant which dates it's showing you.

An example with numbers

Example with made-up numbers: you ask an agent to "raise the budget on campaigns that run out of budget." It finds 12 campaigns at $40 a day and proposes $60. That's $240 more per day, about $7,200 over 30 days if every campaign spends its cap. Approved in one click during peak season, that change can eat a month's margin. Reviewed campaign by campaign, you might approve four.

Official servers and third-party servers

An official server is built and maintained by Amazon: that's the case for the Ads MCP Server and the Selling Partner plugin. A third-party server comes from another company or a developer, and it reaches your account like any other app: you authorize it on the SP-API or the Ads API, and you revoke it the same ways described above.

For a seller, the practical differences come down to trust and control:

  • Who stores your data and what happens to it when you disconnect. With a third-party server there are two privacy policies involved, the server's and the AI assistant's.
  • Which tools it exposes and whether any of them write. The MCP specification says clients must treat a tool's behavior labels (such as "read only") as untrusted unless they come from a trusted server.
  • What permission you gave it. In Ads, authorizing with a view-only user caps what the app can do, official or not.

In its help article on custom connectors, Anthropic notes that it doesn't verify those connectors, and recommends connecting Claude only to servers from organizations you trust, reviewing the permissions they request, and disabling tools that write to external apps if you don't need them (Anthropic, custom connectors).

How we read it at AGTA

This is our own judgment, not something Amazon says.

  • We'd start read-only: in the plugin, limited to answers; in Ads, authorized with a view-only user. That gives you a few weeks to check whether the assistant gets things right before it changes anything.
  • If you grant write access, we'd switch off the delete tools and approve each change to bids, budgets and listings one at a time, never in bulk.
  • Before acting on a number, we ask the assistant for the source and the date range and check it against the Seller Central or Ads console report, because a well-written answer can rest on data that's three days old.
  • To read what the assistant returns, it helps to have the basic metrics down: what ACoS is and how to read it and the Search Query Performance funnel.

Sources