Amazon Ads MCP Server: what it does and how to connect it

In short: The Amazon Ads MCP Server is the interface through which AI agents call Amazon Ads functions. It turns a prompt into structured API calls. One special case: the expert tool `amc_sql_generator` produces validated AMC SQL from a question in natural language. Access requires API credentials, and that tool is in closed beta.

What the server is and what it is not

The Amazon Ads MCP Server has been available in open beta since February 2, 2026. It translates natural language into structured calls to the Amazon Ads API, so an agent can use Amazon Ads functions without building a separate integration for each one. Amazon names Claude, ChatGPT, Gemini, Kiro, Amazon Bedrock and Amazon AgentCore as client examples.

Important for orientation: there is no separate product called an "Amazon Marketing Cloud MCP". The server belongs to the Amazon Ads APIs. Amazon Marketing Cloud is the analytics side of this world, and it is reachable through exactly one tool: the AMC SQL Generator. Anyone searching for an "AMC MCP" finds third-party offers that bridge AMC data through their own connector. That is something else than Amazon's server.

Standard tools and expert tools

Amazon distinguishes two kinds of tools. Standard tools are deterministic actions: change a budget, create a campaign, generate a report, read account information. Expert tools interpret the request internally and return the finished result. The first expert tool is the AMC SQL Generator.

The generator takes a question in natural language and returns a validated AMC SQL query that is ready for execution. It understands the AMC data structure, the audience capabilities and the best practices, and thereby translates requirements that would otherwise need SQL expertise. The result is a query plus an explanation of the logic, so you can review the output before it runs.

Tool typeWhat happensExample from the Amazon docs
Standard toolA concrete action is executed, the result is deterministic.Increase a campaign budget to $500; create a campaign report for an account.
Expert toolThe request is interpreted internally, Amazon intelligence returns the finished result.Build an audience of shoppers who purchased my ASINs in the last year but have not re-purchased in the last 90 days.

Requirements for the connection

For access you need an Amazon developer account and an approved application with access to the Amazon Ads API. The server does not replace that approval, it presupposes it: a prompt alone is not enough to reach advertising data.

For the connection itself Amazon names two ways. The classic way uses a Login with Amazon profile: every request carries the header Amazon-Ads-ClientId for your LwA application and Authorization with the bearer access token. The second way is OAuth 2.1: the MCP client handles the token management itself through a browser login, so no tokens sit in the configuration file. Amazon documents Kiro CLI, Claude Code, Claude Desktop and ChatGPT for this. The required scope is advertising::campaign_management.

  • An Amazon developer account and an approved application with access to the Amazon Ads API; in open beta the server is available globally to partners with active API credentials.
  • A Login with Amazon security profile with a client ID; with OAuth 2.1 you also add the callback URL of your client to the allowed return URLs.
  • The scope advertising::campaign_management; an MCP-capable client program such as Kiro CLI, Claude Code, Claude Desktop or ChatGPT.

Regional endpoints and account context

The endpoints are not the same across regions. Amazon lists NA, EU and FE for the regular version, and for the Lite version the additional path /mcp/lite. The EU endpoint therefore sits in the European region, not in the North American one.

For the account context Amazon distinguishes two modes. In dynamic account context the model asks for the account identifiers itself; that is the default. In fixed account context you set the identifiers statically as headers, among them Amazon-Ads-AccountID, where an AMC instance ID can also be used. The mode is switched on with Amazon-Ads-AI-Account-Selection-Mode: FIXED; in that mode every request additionally needs at least one of the account headers.

RegionEndpoint (regular)Endpoint (Lite)
NA (North America)https://advertising-ai.amazon.com/mcphttps://advertising-ai.amazon.com/mcp/lite
EU (Europe)https://advertising-ai-eu.amazon.com/mcphttps://advertising-ai-eu.amazon.com/mcp/lite
FE (Far East)https://advertising-ai-fe.amazon.com/mcphttps://advertising-ai-fe.amazon.com/mcp/lite

The Lite variant with three tools

The Lite version provides the same Amazon Ads functions, but it does not expose them as a separate tool per operation. Instead your agent gets three meta-tools with which it searches the catalog on demand, reads the schema of a tool and executes it.

That saves context: the full catalog is not loaded into the model in every session, only the parts the current task needs. Amazon describes the Lite variant as functionally equivalent; the only difference is how the tools are presented to your model. For smaller models or sessions with a few operations, Lite is the sensible choice.

  • amazon_ads_mcp-search_tools searches the tool catalog with a question in natural language and returns the matches along with the other tools of the affected area.
  • amazon_ads_mcp-describe_tools retrieves the full input schema for one or more tools.
  • amazon_ads_mcp-invoke_tool invokes the tool with the arguments assembled from that schema.

The AMC SQL Generator: what it delivers and what it does not

The AMC SQL Generator (amc_sql_generator) is the first expert tool and is in closed beta according to the Amazon documentation; access runs through an application form and is reviewed for technical readiness. The path into AMC therefore leads through a single, clearly bounded tool and not through a general AMC access inside the server.

Who benefits? Amazon names tech providers, agencies, sellers, vendors and enterprises with an active AMC instance, a developer account and Amazon Ads API access who are building agentic advertising workflows. The practical gain lies in the review step: you get SQL plus an explanation, check fields and thresholds against the schema reference, and let the query run in the sandbox first before it counts in the instance.

Two limits belong with it. The tool is a generator for queries, not a replacement for the AMC data path: the query still has to run against your AMC instance setup, and the aggregation thresholds of AMC apply unchanged. And anyone who wants to create an audience via MCP still needs the AMC audience path itself, not just a prompt.

A first run in practice

In practice a first run takes four steps. First you sign in with your LwA profile and retrieve the access and refresh tokens, or you use the OAuth 2.1 flow of your client. Second you enter the server with client ID and Authorization header in the configuration of your MCP client. Third you start the client and check the connection, usually with a list command for the available tools or accounts. Fourth you set a concrete task.

For the AMC part, a good first assignment is the formulation, not the execution: "Build an audience of shoppers who purchased my ASINs in the last year but have not re-purchased in the last 90 days." You thereby get a query as a draft that you check against the schema, the thresholds and the rules for user_id before it runs anywhere.

Sources: Amazon Ads API: Amazon Ads MCP Server overview · Amazon Ads API: AMC SQL Generator · Amazon Ads API: Connecting to the Amazon Ads MCP Server · Amazon Ads API: Connecting to the Amazon Ads MCP Server (Lite) · Amazon Ads API: Connecting with OAuth 2.1 · Amazon Ads: Introducing the Amazon Ads MCP Server (retrieved 2026-11-14)

Keep reading: Writing AMC SQL with AI · AMC API migration · AMC sandbox