Amazon Ads runs its own MCP server. It is a hosted endpoint that lets an AI client such as Claude or ChatGPT call the Amazon Ads API in plain language: list your advertising accounts, pull a campaign report, raise a budget, create a Sponsored Products campaign. Amazon announced it in open beta in February 2026 and has kept adding to it since.
This guide explains what the Amazon Ads MCP Server is, who can use it, what its tools and skills do, how authentication works, how to connect it to Claude and ChatGPT, and where it stops. Everything about Amazon's server below was read on Amazon's own documentation, news and release notes pages on 11 November 2026. The last part covers what sits outside an ads-only server and when a data layer next to it helps. If you came here for DataDoe's hosted Amazon Ads MCP server, that has its own page.
MCP, the Model Context Protocol, is an open standard for how AI applications talk to outside tools. Amazon describes its server as a translation layer between the two sides: it connects AI agents to Amazon Ads API functionality and turns natural language prompts into structured API calls. Amazon builds, maintains and updates the server.
Amazon does not own or operate the AI model or the application you connect. You bring the client. The documentation names Claude, ChatGPT, Amazon Kiro, Amazon Bedrock and Amazon AgentCore, and says other MCP-compatible applications work too.
The server is still a beta. Amazon's news post of 2 February 2026 announced the open beta and says the server and its tools are available globally to Amazon Ads partners with active API credentials. The release notes entry of 2 November 2026 still calls it the MCP beta.
In practice, "active API credentials" means three things, all listed as prerequisites in the setup guides:
If you do not have those, the guide sends you to the Amazon Ads API onboarding steps first. For a seller who only wants to chat with an ad account, this is the main hurdle: the server expects a developer setup, not just an advertiser login.
The server is regional. There are three endpoints, and Amazon notes they are not interchangeable:
A tool is a function the server exposes to the agent. Amazon splits its tools into two types.
The regular server registers every operation as its own tool. Amazon's own guide for Amazon Quick says that is more than 100 tools. Since 29 September 2026 there is a Lite version on the /mcp/lite path. It exposes three tools, search_tools, describe_tools and invoke_tool, and the agent works in a loop: search the catalog, read the schema of the tool it picked, then invoke it. Amazon states that Lite is functionally equivalent to the regular server and suits smaller models and cost-sensitive sessions.
Skills are the second building block. Amazon defines a skill as a pre-defined prompt definition that guides an AI application through an Amazon Ads workflow and its tool usage. They are files, not functions. You download them from a separate skills endpoint with the same client ID and access token as the server. For Claude Code, the guide says to unpack them into the .claude/skills folder.
In July 2026 Amazon added three skills: accepting advertising terms before an advertiser's first account is registered, the user invitation lifecycle, and user permissions. In November 2026 it added AI migration skills, which scan a developer's codebase and suggest changes for moving from older campaign management APIs to the current version. So far this is account administration and developer work, not campaign strategy.
The server reads and writes. Amazon's announcement lists four areas an agent can reach once connected:
Later additions include AMC workflow tools, which have covered all AMC instances since May 2026, and three audience discovery tools added on 30 October 2026.
The catalog includes delete tools. Amazon's setup guide shows how to switch individual tools off in the client configuration, using the delete campaign and delete ad tools as its example. If you only want reporting, that is the place to say so.
There are two ways in, and both run through Login with Amazon.
Tokens in the config file. You complete an authorization grant yourself, then send two headers with every request: your client ID and a bearer access token.
OAuth 2.1. Since 6 July 2026 the MCP client can fetch tokens itself through a browser login. You add the client's callback URL to the allowed return URLs of your security profile, add the server to the client, sign in with your Amazon advertising account and allow the permissions. Amazon describes two modes. A public client supplies only the client ID, and its access token expires after one hour, after which a browser popup asks you to sign in again. A private client adds the client secret, stores a refresh token and renews without the popup.
Accounts are handled in one of two modes. By default the model asks which account to use. In fixed mode you set the profile, account or manager account ID as headers, and the server stops asking. Amazon notes that all fixed headers must point at one account.
Amazon documents OAuth setup for Kiro CLI, Claude Code, Claude Desktop and ChatGPT.
It is an Amazon Ads server, and its scope ends where the Ads API ends. Nothing in Amazon's MCP documentation mentions orders, referral and FBA fees, product costs, inventory or Vendor Central. Amazon has a second MCP server for Seller Central, and how the two differ from DataDoe is laid out in our comparison with Amazon's MCP servers.
Three consequences follow for anyone managing ads for profit.
Independent and open-source Amazon Ads MCP servers also exist. They work with the same advertising data, so the same boundary applies.
If you are a developer or an agency with Ads API credentials and your questions stay inside the ad console, Amazon's server may be all you need.
A data layer earns its place when the question crosses that boundary. DataDoe loads Amazon Ads, Seller Central and Vendor Central into one schema and serves it over MCP, a REST API and BigQuery. Your AI then reads ad spend next to orders, Amazon fees and the product costs you upload, in tables such as amazon_ads_profit_by_target_and_date and amazon_ads_profit_by_search_term_and_date. The profit columns are estimates built from your costs and recent settlements.
The data is stored. The daily fetch starts at 2am in the seller's local time, campaign and product ad performance also update during the day, and DataDoe keeps each day it collects from the day you connect. An hourly campaign table supports dayparting analysis.
DataDoe can also change things in Amazon Ads through Actions: campaigns, ad groups, ads, targets and portfolios, budget rules for Sponsored Products, Brands and Display, and scheduled bid increases for Sponsored Products. Actions stay off until you enable them, most AI clients ask before a change, and every change is logged. A dry run validates any Action before it is sent. Amazon DSP is read only.
Setup needs no developer registration. You sign in with Amazon inside DataDoe, then in Claude open Customize → Connectors → Add and paste https://mcp.datadoe.com/mcp/v1.
The pages we read on 11 November 2026 list no price for the server. You need Amazon Ads API access, and you pay for the AI client you connect. Amazon has said expert tools will be pay per use once they are generally available.
Yes. Amazon documents OAuth 2.1 setup for Claude Desktop, Claude Code and ChatGPT. You need the client ID of an approved Login with Amazon application with Amazon Ads API access.
Yes. Its tools create, update and delete campaigns and other entities. Your client configuration can disable the tools you do not want an agent to call.
Ad metrics. It covers advertising data, account settings and billing. Product costs, Amazon fees and orders are outside it, so profit after ads needs a second source.
DataDoe is the Amazon data layer for AI agents. It puts Amazon Ads, Seller Central and Vendor Central into 130+ tables that Claude, ChatGPT and other MCP clients can query and, when you allow it, act on. The price is $97 a month and the first 14 days are a free trial. Connect an Ads profile on the Amazon Ads MCP page, upload your product costs, and ask which search terms made money last month.
.svg)
