MCP for e-commerce: connecting Claude and ChatGPT to Shopify, Seller Central and Amazon Ads

Claude and ChatGPT are good at reading a report you paste in. They become far more useful when they can look up your orders, stock and ad performance themselves. That is what an MCP server for business does: it gives an AI assistant a defined, limited set of ways to read, and sometimes change, data in the systems you already use. In 2026, MCP for ecommerce stopped being a developer experiment: Amazon and Shopify now ship official servers and plugins of their own.

This guide is for store owners, marketplace sellers and their operations teams. It covers what an MCP server is in plain English, what is official today, read-only versus write tools, approvals, and when a private server over your own data makes sense. Platform facts are as of October 2026.

I build AI agents and marketplace automation, and my own site runs a small read-only MCP server, which I use as an example below.

What an MCP server is, in plain English

The Model Context Protocol (MCP) is an open standard, introduced by Anthropic in late 2024, for connecting AI assistants to tools and data. In December 2025 Anthropic donated MCP to the Agentic AI Foundation, a fund under the Linux Foundation, noting more than 10,000 active public MCP servers and adoption by ChatGPT, Gemini, Microsoft Copilot and others.

The plain-English version: an MCP server is a menu of labelled actions an assistant is allowed to use. "Get yesterday's orders." "Get stock for this SKU." "Draft a price change." The server decides what is on the menu and what each item does behind the scenes. The assistant decides when to ask for one. You decide which servers the assistant can reach at all.

Three terms are worth knowing:

• Tools are actions the assistant can call, such as looking up an order.
• Resources are data it can read, such as your returns policy.
• The host is the app you chat in, such as Claude or ChatGPT. It connects to the server and should ask your permission before using it.

How it differs from an API: your store's API is a door for software. An MCP server stands in front of that door and describes a few chosen actions so that any MCP-capable assistant can use them.

What is official today: Shopify, Amazon Ads and Seller Central

Three official routes matter for most sellers.

Amazon Ads MCP server. Amazon Ads opened it in open beta on 2 February 2026, globally, for Amazon Ads partners with active API credentials. It turns natural-language requests into Ads API calls and works with Claude, ChatGPT, Gemini and custom agents. It can read (reporting, billing) and write (creating, updating or deleting campaigns), and one tool builds a complete Sponsored Products campaign that ends ready for review and approval. Advertisers without their own API credentials reach it through a partner or developer who has them.

Shopify's agent tools. Shopify's documentation for agents lists catalogue search across all Shopify merchants or within one store, Cart and Checkout MCP servers, an Order MCP with a get_order tool, and order webhooks. These mainly serve shopping agents that buy from stores, the side I cover in agentic commerce for online stores. For running your own store, Shopify's Sidekick assistant works inside the admin, and anything more custom goes through the Admin API.

Amazon's Selling Partner plugin for Claude. Announced at Amazon Accelerate in September 2026, it lets sellers connect Seller Central data to outside AI tools, starting with Claude and Amazon Quick, and starting in the US. Amazon's Accelerate 2026 recap describes it as bringing "selling data and curated skills" into those tools, such as reading sales trends and working on listings.

None of these joins things up: each official route covers one platform. If you also sell on Etsy or TikTok Shop, keep costs in a spreadsheet, or want one answer that combines channels, that still needs custom work.

Read-only tools and write tools

The most important design question is not which assistant you use. It is whether each tool reads or writes.

	Read-only tools	Write tools
Examples	Sales by SKU, stock levels, order status, ad reports	Change a price, pause a campaign, edit a listing, refund an order
Worst case	A wrong answer you can check	A wrong change on a live account
Sensible default	Allowed and logged	Drafted, then approved by a person

MCP lets a server label a tool as read-only, but the specification also warns that such descriptions "should be considered untrusted, unless obtained from a trusted server". So do not rely on labels. Rely on credentials: a read-only server should hold API access that cannot write at all.

When you do add write tools:

• Separate them. A different server, or at least different credentials, from the read tools.
• Put limits inside the tool. A maximum bid change, a price floor, a daily cap on edits. The assistant cannot talk its way past a limit in code.
• Draft first. The tool prepares the change; a person applies it.
• Log every call. Who asked, what changed, the value before and after.

Approval, safety and Amazon's Agent Policy

The MCP specification says hosts "must obtain explicit user consent before invoking any tool". That is a good floor, but do not rely on someone reading every prompt carefully. Approval for anything that changes money, stock or ads belongs in your own system, where it is logged.

Amazon adds rules of its own. Its Seller Forums announcement says that from 4 March 2026 the Business Solutions Agreement includes a new Agent Policy, and that all AI agents must "clearly identify themselves as automated systems", "comply with the new Agent Policy at all times" and "cease access if Amazon requests". For an MCP setup on an Amazon account, that means:

• Use the official APIs (Selling Partner API, Amazon Ads API), not browser automation that imitates a person.
• Make the tool identifiable as automated.
• Build an off switch you can use in seconds, and honour Amazon's requests to stop.
• Keep the audit log described above.

I go into the implementation in keeping automation compliant with Amazon's Agent Policy.

An example: the read-only MCP server on devwren.com

My own site runs a small MCP server at https://devwren.com/mcp. It has seven read-only tools: profile and contact details, services, plans and starting prices, projects, blog guides, a list of pages, and reading any page in full. Every answer comes from files the site already publishes, so it can only say what the site says. It cannot send messages or book work; when someone wants to start a project, it hands back the contact details. It is documented, with the site's other machine-readable endpoints, on the developers page.

It is small, but it shows the pattern I recommend for a first store server: a narrow set of read tools, answers from one source of truth, no ability to act, and a clear route to a person.

Building a private MCP server for business data

A private MCP server for business data makes sense where the official ones stop: several channels, data in spreadsheets or an ERP, or questions that combine sources, such as "which SKUs are selling fast on Shopify but running low at FBA?"

1. List the questions. Five to ten that you or your team ask every week. They become the tools.
2. Connect the official APIs. Shopify's Admin API, Amazon's Selling Partner and Ads APIs, Etsy's Open API v3, TikTok Shop's Open API. One connection per channel, with the accounts in your name.
3. Keep a copy of the data. Sync it into your own database on a schedule or from webhooks, so the assistant's questions do not run into marketplace rate limits and answers stay quick.
4. Decide who can connect. A sign-in for each person, and tools scoped to their role.
5. Log everything. Every call, its inputs and its result.
6. Add write tools last. One at a time, as drafts a person approves.

The current MCP specification (2026-07-28) uses stateless HTTP, so a remote server can be a small web service rather than something running on one person's laptop.

What it costs: official tools or a server of your own

• Official servers and plugins come from the platforms, but each covers one platform and needs its own access, such as Amazon Ads API credentials through a partner.
• A private MCP server is custom AI agent work. I quote it from a short brief rather than a price list, because the scope varies widely. What moves the quote: the number of channels, whether you need write tools and approvals, how many people connect and how much history you keep.
• If most of the job is data plumbing, connecting a marketplace through its official API and keeping the data in sync, it overlaps with my marketplace plans on the pricing page, which start from $2,500 for an MVP with one marketplace connected. That is a starting price; every project gets a fixed quote after a brief. The e-commerce automation page shows the channels I work with.

FAQ

What is the difference between an MCP server and an API?

An API is how software talks to a platform such as Shopify or Amazon: precise, technical and different for every platform. An MCP server sits in front of one or more APIs and offers an AI assistant a short menu of named actions, described in a standard way that Claude, ChatGPT and other MCP clients understand. You still need the API underneath; MCP makes it usable by an assistant without a separate plugin for each.

Is it safe to let Claude or ChatGPT make changes to my Amazon or Shopify account?

It can be, if the server is built for it. Start read-only, with credentials that cannot write. When you add write tools, have them prepare drafts that a person approves, put hard limits inside each tool, log every call and keep an off switch. On Amazon, the Agent Policy in force since 4 March 2026 also requires agents to identify themselves and stop when Amazon asks.

Can I connect Claude to Seller Central without a developer?

For Amazon data alone, increasingly yes: Amazon announced its Selling Partner plugin for Claude in September 2026, starting in the US. You need a developer when you want several channels in one place, data from your own spreadsheets or systems, tools shaped around your process, or write actions with approvals and limits. For many sellers, the official plugin is the right first step.

If you want Claude or ChatGPT working with your store data, send me a short brief: which channels you sell on, the questions you ask every week, and whether the assistant should only read or eventually make changes. You can send it through the contact form, see the AI agents for business service, or read my guide to what AI agents can do for a business.