Shopify MCP in 2026: Which Server Does What, and How to Connect Claude

Shopify runs several separate MCP servers, and each one serves a different job. Dev MCP gives coding assistants Shopify's docs and GraphQL schemas. The Catalog, Cart, Checkout and Order servers let shopping agents find and buy products under the Universal Commerce Protocol (UCP). Customer Accounts MCP handles signed-in buyer requests such as order status. None of them is a merchant analytics server. If you want Claude or ChatGPT to answer "what did returning customers spend last month after the Meta campaign", you need a data source that joins Shopify with ads and email, which is what Finsi MCP is built for.

Below is what each server does, how to connect Claude to it, and where it stops. Everything here comes from shopify.dev and Shopify's npm packages as of November 2026. Shopify has changed this area several times this year, so check the linked docs before you build on it.

The Shopify MCP servers at a glance

Shopify also ships the Shopify AI Toolkit, a plugin for Claude Code, Codex, Cursor, VS Code and other tools. It bundles skills for docs lookup, GraphQL validation and running store tasks through Shopify CLI.

Dev MCP: Shopify docs inside your editor

Most "Shopify MCP server" searches are really about Dev MCP. It runs locally, can't see any store, and exists so a coding assistant stops guessing field names in the Admin API. Its README lists two tools: search_dev_docs and introspect_admin_schema. There's also a prompt for writing Admin GraphQL operations against the current schema. When we checked on November 5, 2026, the npm package was at version 1.18.0, so the list may be longer now.

To add it to Claude Desktop or Cursor, put this in the MCP config file:

{
  "mcpServers": {
    "shopify-dev-mcp": {
      "command": "npx",
      "args": ["-y", "@shopify/dev-mcp@latest"]
    }
  }
}

In Claude Code, Shopify now points people to the AI Toolkit plugin instead:

claude plugin install shopify-ai-toolkit@claude-plugins-official

For Codex the equivalent is codex plugin add shopify@openai-curated. Shopify's docs list Node.js 22.19.0 or newer as a requirement.

Catalog, Cart and Checkout: MCP for shopping agents

The old Storefront MCP endpoint at https://{shop}/api/mcp lost its catalog and cart tools. Shopify's migration page sends developers to UCP endpoints at https://{shop}/api/ucp/mcp instead. The tool map from that page:

Two details trip people up. Every request to /api/ucp/mcp has to include an agent profile URL (meta.ucp-agent.profile), and the tools you get back depend on the capabilities that profile declares. And each update_cart call replaces the whole cart, so you resend every line item you want to keep.

Global Catalog search returns results grouped by a Universal Product ID with offers from several merchants. Some fields, such as description and metadata.top_features, are generated or enriched by Shopify and tagged Inferred in the reference. That matters for brands. What a shopping agent says about your product may come partly from Shopify's inferred fields rather than your own copy. Shopify's Fall '26 Editions added a place in the admin to review and approve the business facts, policies and FAQs that get shared with agents. If you sell on Shopify, that screen is worth ten minutes this quarter.

Customer Accounts MCP: orders for a signed-in buyer

This server answers buyer questions such as "where is my order" for an authenticated customer. You don't get a fixed URL. The app fetches https://{shopDomain}/.well-known/customer-account-api and reads the mcp_api field from the response. Shopify lists three requirements: the store needs a custom domain, the app has to meet Shopify's protected customer data requirements, and requests carry an OAuth 2.0 token obtained through the authorization code flow with PKCE.

It's a support and post-purchase tool. It doesn't give a merchant a view across customers.

What about Admin data and store changes?

Shopify doesn't document an Admin MCP server for merchants. The closest official path is the AI Toolkit's store execution skill, which has the assistant write and validate an Admin GraphQL operation and then run it through Shopify CLI:

shopify store auth --store your-store.myshopify.com --scopes read_products
shopify store execute --store your-store.myshopify.com --query '{ products(first: 10) { nodes { title } } }'

The skill asks for the narrowest scope the operation needs, and a mutation only runs when the command includes --allow-mutations. That flag is the approval step, and it lives in a developer's terminal. For a brand team that wants an assistant to suggest price or inventory changes, the questions are who approves, where the record of the change lives, and what happens to the CLI session when that developer leaves.

Third-party "Shopify MCP" servers on GitHub fill the Admin gap, usually with a personal access token. Read the code first. Then read the scopes it asks for, and only then connect it to a live store, because a token created for a weekend experiment tends to keep working long after anyone remembers it.

Where analytics questions fit

None of the servers in that table answers business questions across channels. Inside the admin, Shopify's answer is Sidekick, which can build ShopifyQL reports and customer segments from store data. We compare it with outside agents in Shopify Sidekick vs AI agents.

When the question mixes Shopify orders with Meta spend, Klaviyo flows or Recharge subscriptions, the model needs those sources joined and defined in one place first. Otherwise Claude does arithmetic on CSV exports and picks its own definition of "new customer".

That is the gap Finsi MCP covers. Finsi loads Shopify, Klaviyo, Meta Ads, Google Ads, Recharge and other sources into one warehouse, defines metrics once in a semantic layer, and exposes them to MCP clients as read-only tools such as get_revenue_summary, get_ads_performance, get_customer_cohorts and run_catalog_metric. Each API key covers one brand, every call is logged, and the MCP server can't change ads, flows, listings or prices. Changes go through approval inside Finsi. The full tool list is on the Finsi MCP page, and the operating model is in Bring your own AI agents to Shopify data via MCP.

Adding it to Claude Code is one command:

claude mcp add --transport http finsi https://api.finsi.ai/mcp \
  --header "Authorization: Bearer finsi_<your_key>"

A sensible setup for most brands runs both. Shopify's Dev MCP or AI Toolkit for whoever edits the theme and apps, and Finsi MCP for the people asking money questions. Neither needs a personal Admin token.

FAQ

Does Shopify have an official MCP server?

Yes, several. Dev MCP for developers, UCP-based Catalog, Cart, Checkout and Order servers for shopping agents, and Customer Accounts MCP for signed-in buyers. They are separate products with separate endpoints.

What is the Shopify Dev MCP server?

A local MCP server, installed with npx -y @shopify/dev-mcp@latest, that lets coding assistants search shopify.dev and introspect the Admin GraphQL schema. It has no access to store data.

How do I connect Claude to Shopify?

It depends on the job. Developers should install the Shopify AI Toolkit plugin in Claude Code, or add Dev MCP to Claude Desktop's config file. Store tasks go through shopify store execute. Revenue, cohort and ad questions across your stack need a metrics source such as Finsi MCP.

Is Storefront MCP deprecated?

Its catalog and cart tools on /api/mcp were removed. Shopify now serves those capabilities through UCP at /api/ucp/mcp, and every request needs an agent profile. Customer Accounts MCP was not changed.

Can an MCP server change my Shopify store?

Shopify's shopping MCP servers handle carts and checkouts for buyers. They have no access to your admin. Store changes made through the AI Toolkit need an authenticated CLI session plus the --allow-mutations flag on every write. Finsi MCP is read-only, and changes proposed by Finsi agents wait for a person to approve them.

Related

• Finsi MCP
• AI agents on Finsi
• Approval-gated AI agents for ecommerce
• Shopify analytics beyond the built-in reports