Shopify with Claude or ChatGPT
Three ways to plug a store into an AI assistant, what each one can read and change, and the one that takes four minutes.
You want your assistant to answer questions about your store with the store's own data: which products have thin descriptions, how last week's orders compare with the week before, what the live theme actually runs. There are three ways to wire that up, and they differ on three things: who holds the pen, who holds the keys, and how much code you end up owning.
This guide compares them. The step-by-step for each client lives in its own guide, one per client, and this page links to them rather than repeating them.
The options
Three ways in
| | Custom connector (Shopify MCP) | Shopify's own connector | Your own code | |---|---|---|---| | What you do | paste one URL, sign in, approve scopes | add it from Claude's connector directory | call the Admin API, hand results to the model | | Clients | Claude, ChatGPT, Claude Code, Cursor | Claude | the one you build | | Reads the store | yes, within your Custom App's grant | yes | what you code | | Writes to the store | no, refused before Shopify | yes | what you code | | Code to maintain | none | none | all of it |
1. A remote MCP server as a custom connector. The store gets one URL. You paste it in your assistant, sign in, and approve what it may read. BoostEcom's Shopify MCP works this way, and so does any remote MCP server.
2. Shopify's own connector. Shopify publishes a connector in Claude's directory, and it writes to the admin: products, inventory, discounts. If you want the assistant to change the store from a conversation, that is the tool built for it. The two are compared line by line in Shopify's official Claude connector or a read-only relay.
3. Your own code. A Shopify Custom App gives you an Admin API token; your code queries the store and passes the results to the model through tool use (Anthropic) or function calling (OpenAI). Full control, and every retry, rate limit and model upgrade is yours to maintain.
The protocol
What MCP changes
The Model Context Protocol is an open standard for exactly this job. A server announces its tools, with typed arguments, when a client connects; the client calls them by name. You write no tool loop and install no SDK, and the same URL serves every client that speaks MCP over HTTP.
That is why paths 1 and 2 need no code: the assistant already is the MCP
client. And it is why path 3 has shrunk. Both vendors' APIs can now call
a remote MCP server themselves: Anthropic's Messages API through its MCP
connector, which takes the server URL and an authorization_token, and
OpenAI's Responses API, which accepts a remote MCP server as a tool. You
keep the orchestration and drop the glue.
Claude
Connecting Claude
Claude takes a remote MCP server as a custom connector: in Claude's settings, open Connectors and add a custom connector. Paste the store's URL and leave the client id and secret empty. Claude identifies itself, runs an OAuth 2.1 handshake with PKCE, and opens a consent screen that names the client, the store and every permission it asks for. Approve, and Claude's tool list fills with the tools those permissions open.
→ Connect Claude, step by step
Claude Code signs in the same way from a terminal: claude mcp add
declares the server, then /mcp opens your browser on the same consent
screen. → Connect Claude Code
ChatGPT
Connecting ChatGPT
ChatGPT takes a remote MCP server once Developer mode is on: in ChatGPT's settings, turn on Developer mode, then add a connector with the store's URL and OAuth. Developer mode is available on the Plus, Pro, Business, Enterprise and Edu plans. ChatGPT asks you to confirm before a tool runs, and you can keep that answer for the rest of the conversation.
→ Connect ChatGPT, step by step
Cursor reads the same URL from its mcp.json and signs in from its MCP
settings. → Connect Cursor
Choosing
Claude or ChatGPT?
For this job, the relay does not care: both receive the same tools under the same grant, and neither runs a model on our side. Pick the assistant your team already uses, on whichever model your plan offers. What actually differs is how you get connected:
| | Claude | ChatGPT | Claude Code | Cursor |
|---|---|---|---|---|
| Where you add it | Settings, Connectors | Developer mode, Connectors | claude mcp add, then /mcp | mcp.json, then MCP settings |
| Sign-in | OAuth | OAuth | OAuth | OAuth |
| Static key for CI | not needed | not needed | yes | yes |
| Guide | Claude | ChatGPT | Claude Code | Cursor |
Reach
What the Shopify MCP reads
Every connected store gets its own endpoint:
https://www.boostecom.app/api/mcp/v1/<STORE_ID>
That endpoint registers 24 tools. 22 of them can be offered to a client, and the other 2 are reserved to BoostEcom admins: they never appear on a customer's consent screen. They are 15 Shopify relay tools plus BoostEcom's own read tools (the Studio, the store's intelligence record and the documentation). The tools and scopes reference lists every one, with the scope that opens it.
Two gates decide what a client sees. The scopes you approve on the
consent screen, and the scopes Shopify granted your Custom App: a store
connected with read_products alone never shows the order or theme
tools, whatever you approve. A static key, used for CI and headless runs,
never passed a consent screen, so it opens the relay tools and nothing
else.
The relay returns Shopify's own query cost and throttle status beside the data, so the model can pace itself instead of guessing.
Writes
Read-only, on purpose
Through the Shopify MCP, neither Claude nor ChatGPT changes your store. Queries run; every mutation, and any GraphQL document the relay cannot prove only reads, is refused before it reaches Shopify, and the attempt is written to your audit log.
A write to a live store is a decision someone has to approve, and a relay has no human in it. So changes go elsewhere: through Shopify's own connector if you want the assistant to hold the pen, or through BoostEcom's dashboard chat, where @Atlas proposes and you approve each change.
Build it yourself
If you write your own integration
Path 3 is the right call when you need a workflow nobody ships, or data that must never leave your own infrastructure. Three habits keep it safe:
- Scope the Custom App like a new hire. Start with read scopes
(
read_products,read_orders,read_content,read_themes), watch what the model asks for for a week, then add a write scope for one surface at a time. - Never give a model the scopes a mistake cannot undo: payment settings, staff accounts, files.
- Check every tool call before it reaches Shopify: the tool, the scope it needs and the store it targets. A model can emit a well-formed call with the wrong store id.
Cost
What it costs
The Shopify MCP is free on every BoostEcom plan, Free included. It forwards a Shopify Admin API call and runs no model, so it spends no credits; per-organization rate limits exist to stop abuse, and the Shopify MCP page lists them per plan.
Your Claude or ChatGPT subscription is separate and unchanged: the assistant runs on your own account. BoostEcom's own chat, where @Atlas and the specialists work on the store, is what spends credits.
FAQ
Frequently asked questions
Is there a one-click install?
No. You connect the store to BoostEcom once, through a Shopify Custom App (its client id and secret go in the store's settings, under Connectors), then paste one URL in your assistant. About five minutes the first time.
Do I need Shopify Plus?
No. Custom Apps and the Admin API are available on every Shopify plan.
Does the assistant ever see my Shopify token?
No. The client signs in to BoostEcom, never to Shopify, and only ever holds a BoostEcom token. Revoke it and the client is cut off, with nothing to change in Shopify.
Do I have to host anything?
No. BoostEcom hosts the server, and it keeps no session between calls, which is why one URL serves every client.
Which model does it use?
Whichever one you pick in Claude or ChatGPT. The relay answers tool calls; it runs no model of its own.
Can I bring my own Anthropic or OpenAI API key?
You do not need one for the relay: your assistant uses your own Claude or ChatGPT account. BoostEcom's own chat runs through our AI Gateway accounts and is metered in your organization's credits; it does not take a key of yours.
Pick your client, connect it in four minutes
Each guide gives the URL, the sign-in and what tends to bite, for one client.