Shopify Claude Code: Connector, MCP and How We Use Them

Shopify Claude Code is really three different setups, and mixing them up is how people end up with the wrong tool. The first is Claude, the chat app, connected to your store: you ask, it reads and, if you allow it, writes. The second is Claude Code, Anthropic’s coding agent, used to build themes, apps and Functions on Shopify. The third is agents built on Claude that work on your store with a person approving each step. This guide takes them in that order, with the permissions of each, checked against Anthropic’s and Shopify’s own documentation as of 29 October 2026.

We are noprob agency and we sell the Shopify Launch Sprint, so here is where we stand. We work every day with agents built on Claude Code: the Custom AI System lists development with AI agents among its six superpowers, and this site is built and maintained that way. What follows keeps two things apart: what the documentation says (linked) and the rules we apply (marked as ours). Tool names and limits move fast, so check the linked pages before you rely on a detail.

Which of the three setups do you need?

You need the connector if you want to ask questions about your store, Claude Code if you build on Shopify, and an agent if the same task comes back every week. That is the whole map.

• Claude app with the Shopify connector. For founders and store managers. You talk to your store in plain language: reports, catalog, inventory, discounts. It needs a staff account and your approval in the Shopify admin.
• Claude Code with the Shopify AI Toolkit. For developers and agencies. It works inside your code folder: Liquid theme, apps, Functions, checkout extensions. It needs Node.js, a terminal and a development store.
• An agent built on Claude. For teams with a repeated job, like checking a list every morning. It needs a defined task, narrow tools and an approval step. It costs design time up front.

None of the three replaces Shopify’s own AI: the connector, says Shopify’s help center, does not replace your admin or Sidekick. Start with what the admin already includes, which we cover in Shopify Magic and Sidekick. Our order is always the same: what Shopify includes, then Claude on top, then an agent for the gap.

Can you give Claude access to your Shopify store?

Yes: Shopify publishes an official connector for Claude, and you approve what it can access in your Shopify admin. It is listed in Claude’s directory, made by Shopify and marked as verified by Anthropic. Per Shopify’s guide it is installed from Claude’s directory, not from the Shopify App Store. The setup, as Shopify describes it:

• Open Claude on a desktop browser (mobile availability depends on Claude) and ask for something about Shopify, naming it in the request. Check that the connector is on for the conversation.
• If it is not installed, add it from the connector page in Claude’s directory.
• Claude sends you to your Shopify admin to approve the data access the connector asks for. Shopify makes you confirm in two separate steps.
• Choose the access level. We choose read-only first, for the reason explained below.
• Ask a first question. To work on another store, ask Claude to switch store and authorize it again, then start a new conversation, because an old one can keep using the previous store.

The connector can also create a store before you have a Shopify account: you describe the business and it builds one on the Canvas theme with sample content, password protected so nobody can buy. Shopify warns that anyone with the address of an unsaved store can view and edit it. Treat it as a draft and review the sample content before you sell anything. On cost, Shopify says the connector adds no Shopify fee: standard plan pricing applies, and your Claude subscription goes to Anthropic.

What can Claude do in your store, and what can’t it?

With the connector Claude can read orders, customers and analytics, and edit products, collections, inventory and unpublished themes, but it cannot touch payments or your live theme. This is the split in Shopify’s guide.

• Read only: orders (recent orders or a single one), customers, and analytics through ShopifyQL-based queries. Shopify suggests short questions, one at a time.
• Read and write: products (create, update, search, set draft, active or archived), collections, inventory by variant and location, and percentage-off discount codes.
• Write with your confirmation: pages and menus, publishing an unpublished theme (after your confirmation), and discount codes, where you confirm the start date and the eligible customers.
• Themes: it edits unpublished themes only. To change your storefront, you duplicate the theme in the admin, let the tool edit the copy and publish it yourself.
• Cannot do, even with write access: change your plan, payments, taxes or domains; edit, publish or delete the live theme; refund, cancel or capture orders, mark them paid or process returns; create or adjust gift cards; edit existing customer records or existing discount codes.

Look at the design: money and the storefront stay with you, and the writes that remain have a check built in. Inventory updates read the current quantity first and stop if the count has changed. We copy that pattern in our own agents (more below). For the Shopify side of the protocol, read Shopify MCP explained.

What should you ask Claude about your store first?

Start with questions you can check against the admin in five minutes, one at a time, in read-only. These are ours, built on what the connector reads:

• "Compare sales by channel for the last 30 days with the 30 days before." (analytics)
• "List the draft products that have no description." (catalog)
• "Which variants have the lowest stock in each location?" (inventory)
• "Show my most recent orders grouped by shipping country." (orders)

My rule: the first time, check every answer against the Shopify admin. A report you cannot verify is an opinion. After a week you know which questions Claude answers reliably, and only then does write access for one named task make sense.

How do you set permissions so nothing goes wrong?

Give Claude the least access that does the job and start read-only, because Shopify cannot narrow an approved level: you uninstall the integration and approve again. The rules are on Shopify’s page Authorizing AI tool access to your store:

• Access follows the person. Staff authorize according to their own admin permissions, and a tool never does more than the user who connects it. Our practice: connect with a staff account whose permissions fit the task, not with the owner.
• Two levels. Read-only lets the tool view data; write lets it edit and act on your behalf. If a tool later asks for a higher level, you are prompted and you can decline.
• Use the official integration. Shopify says not to create an API access token or build your own integration for these tools.
• Data leaves Shopify. Once shared with the AI tool it is governed by that provider’s terms and privacy policy. Decide what customer data you put in front of it.
• Revoking is uninstalling. You can remove the integration from the admin at any time.

One more point that is our reading, not a Shopify rule. A store is full of text written by strangers: reviews, order notes, customer messages. Anthropic’s documentation warns that servers which fetch external content can expose you to prompt injection, meaning instructions hidden in the text. So write access and customer-written text should not meet in the same session. We go deeper in AI agent security on Shopify.

How do you use Claude Code to build on Shopify?

Claude Code is Anthropic’s coding agent: it reads your codebase, edits files and runs commands, and the Shopify AI Toolkit gives it Shopify’s current documentation and schemas. Anthropic’s overview describes it as available in the terminal, in IDEs, in a desktop app and in the browser. Shopify’s AI Toolkit page says the toolkit does three things for an AI coding tool:

• Docs and API schemas: it looks up the right Shopify API surface instead of relying on the model’s memory.
• Code validation: it checks GraphQL queries, Liquid templates and Shopify extensions against Shopify’s schemas.
• Store management: it prepares and runs supported store tasks through the Shopify CLI’s authenticated store context, and you decide when a task runs.

The requirement on that page is Node.js 18 or higher. There are three ways to install it in Claude Code, all from Shopify’s documentation:

• Plugin (recommended): `claude plugin install shopify-ai-toolkit@claude-plugins-official`. It bundles everything and updates itself.
• Agent skills: `npx skills add Shopify/shopify-ai-toolkit`, for when you want to manage skills yourself. These do not update automatically.
• Dev MCP server: `claude mcp add --transport stdio shopify-dev-mcp -- npx -y @shopify/dev-mcp@latest`, then restart. Shopify says it runs locally and needs no authentication.

MCP, the Model Context Protocol, is the open standard that connects AI tools to outside data and tools, and Claude Code adds servers with `claude mcp add`. The Dev MCP server needs no login, so by itself it opens no store: it is the documentation and validation side. The store side goes through the Shopify CLI, which Shopify’s help center presents as the connector for developers on their own machine.

What is the difference between the Claude connector and the AI Toolkit?

The connector runs a store from a chat, and the AI Toolkit builds on Shopify from your code folder; they answer different questions. Side by side:

The last line is the one to remember. If you give Claude Code a store through the CLI with write access, nobody has pre-removed the refund button from its reach. Your own rules are the guardrail.

What do we hand to Claude Code on a Shopify project?

We hand Claude Code the work that is repeatable and checkable: Liquid sections, Functions, extension migrations and tests, always on a development store and always reviewed by a person. Three examples that Shopify’s own pages support:

• Theme code. Liquid sections and snippets, written and validated against Shopify’s schemas. In our Custom AI System the theme code is written from scratch for your store only, with no dead code and no app slowing it down, and agents are part of how we develop it.
• Shopify Functions. Functions customize Shopify’s backend logic: discounts, payment and delivery options, checkout validation. Shopify notes that custom apps using Function APIs require a Shopify Plus plan, while public apps from the App Store work on any plan.
• Extension migrations. Shopify’s toolkit page uses this as its example: moving checkout and customer account extensions built on API versions before 2025-10 to Polaris web components, with a prompt as plain as "Migrate the extensions/my-checkout-extension extension to API version 2026-04". Shopify adds that you still review and test the result.

Our rules sit on top of how Claude Code behaves by default. In its permission modes the default one asks before edits, shell commands and network access, and plan mode blocks edits until you approve a plan. So: plan mode first on anything that touches checkout; a development store, which Shopify describes as for development and testing only; the live store receives code through a reviewed change and a duplicated theme you publish yourself. Anthropic’s security page says who decides: you review proposed code and commands before approval.

How do agents on Claude work on a store with human approval?

An agent on Claude is a loop in which the model reads the situation, picks a tool, looks at the result and decides the next step; on a store, the safe version has a person on every write. Claude Code is one such agent. If you want your own, Anthropic’s overview says the Agent SDK lets you build agents on Claude Code’s tools with full control over orchestration, tool access and permissions, and that Claude Code can run on a schedule. What you build on top is the design, and this is ours:

• Read tools by default. The agent lists, searches and compares. Reading is always allowed; writing is a separate decision.
• Writes go through a proposal. The agent states what it wants to change and why, a person approves, and only then does a narrow tool execute.
• Narrow tools. "Set this product to draft" is a tool you can approve. "Run any mutation" is a blank check. The connector itself asks for confirmation before changes made through its GraphQL tools, and we go one step narrower.
• A trail. Every proposal, approval and action is logged, so you can answer "who changed this and why" a month later.
• Start where mistakes are cheap. Drafts, tags, product copy, internal reports. Prices, refunds and customer messages come last, if at all.

The full picture, from chatbot to workflow to agent, is in AI agents for Shopify. A related read: Claude Duet for a Shopify store.

In migrations this is where agents earn their place: repeated checks that a person then approves. We describe the method in migrating to Shopify with AI agents, and the complete path is in our Shopify migration guide. On Nazzi, the redirect work covered 215 archived WooCommerce URLs, and the delivery was the CSV of 301 redirects (case study). That is the kind of work where an agent checks and a person signs off. On Cumini the store came from a proprietary platform tied to the ERP, and the solution was integrated with the ERP (case study).

Checklist: connect Claude to a live store safely

Six steps, in this order, keep a Claude connection useful and reversible. Print it, or paste it into your team’s notes.

• Name the job. One sentence: what do you want Claude to answer or change, and who checks it?
• Pick the account. Connect with a staff account whose permissions match the job, not the owner account.
• Start read-only. Remember that narrowing later means uninstalling and approving again.
• Ask one question at a time and check each answer against the admin for the first week.
• Write access for one named task, and read every confirmation. Theme changes only on a duplicate that you publish yourself.
• Keep customer text away from write sessions, and uninstall the integration when the job is done.

If your store is live and you want AI working on it with a plan, the AI Accelerator is 30 days one-to-one with an established eCommerce, at a fixed price and with no lock-in. If you are still on another platform, moving comes first.

Frequently asked questions

Can you give Claude access to Shopify?

Yes. Shopify’s official connector installs from Claude’s directory, and you approve read-only or write access in your Shopify admin.

Can Claude build a Shopify website for me?

It can start one: the connector creates a store from a description, and Claude Code can write code for a store you develop. You review everything before customers see it.

How do you use Claude to create a Shopify store?

On a desktop browser, name Shopify in your request, describe the business, review the store through the view link, then save it. Treat the first version as a draft.

Is the Shopify connector for Claude free?

Shopify says it has no extra Shopify fee: standard Shopify plan pricing applies, and your Claude subscription costs go to Anthropic.

What is the difference between the connector and Claude Code with the AI Toolkit?

The connector runs a store from a chat. Claude Code with the Shopify AI Toolkit builds on Shopify from your code folder.

Can Claude refund orders or change my live theme?

Not through the connector, per Shopify’s guide, even with write access. With Claude Code the limit is the access you approve, so keep it narrow.

The Shopify Launch Sprint moves your store to Shopify in 14 days with our Custom AI System: a fixed calendar, a public price and 4 written guarantees.

Antonio Manitta

Founder & eCommerce Manager at NoProb Agency

For over 7 years I have helped fashion, supplements, and DTC brands scale their eCommerce. I work in long-term partnership with three brands at a time, building stable systems that compound.

About NoProb →