• A Shopify store, any plan.
• An OpenAI API key (pay-as-you-go).
• Optionally: a ChatGPT Plus, Team or Enterprise seat for Custom GPTs.
• 30 minutes for function calling. 10 minutes for MCP via Responses API. 60 minutes for a polished Custom GPT.

Function calling is the original tool-use API. You declare your Shopify operations as JSON schemas; ChatGPT emits structured calls; your backend executes them.

Step 1 — Create a Shopify Custom App

Same as for Claude. Settings → Apps and sales channels → Develop apps → Create app. Grant scopes by use case (start with read-only).

Step 2 — Wire OpenAI

import OpenAI from "openai"

const openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY })

const tools = [
  {
    type: "function",
    function: {
      name: "get_products",
      description: "List products from the connected Shopify store.",
      parameters: {
        type: "object",
        properties: {
          first: { type: "number" },
          query: { type: "string", description: "Optional Shopify search query" },
        },
        required: ["first"],
      },
    },
  },
] satisfies OpenAI.ChatCompletionTool[]

let messages: OpenAI.ChatCompletionMessageParam[] = [
  { role: "system", content: "You are a Shopify ops analyst." },
  { role: "user", content: "Audit my catalog." },
]

while (true) {
  const completion = await openai.chat.completions.create({
    model: "gpt-5.4",
    messages,
    tools,
  })
  const msg = completion.choices[0].message
  messages.push(msg)
  if (!msg.tool_calls) break

  for (const call of msg.tool_calls) {
    const args = JSON.parse(call.function.arguments)
    const result = await callShopify(call.function.name, args)
    messages.push({
      role: "tool",
      tool_call_id: call.id,
      content: JSON.stringify(result),
    })
  }
}

The loop runs until ChatGPT returns plain text. Standard tool-use pattern, identical in shape to Claude's, only the field names differ.

Pros + cons

| Pro | Con | |---|---| | No extra server | Locked into OpenAI's tool format | | Works on any environment | Pre-declare every tool you need | | Simple to debug | No streaming intermediate state |

OpenAI's Responses API supports MCP servers natively (since 2025 GA). Same MCP server you'd use for Claude. Same tool definitions. One integration, both LLMs.

import OpenAI from "openai"

const openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY })

const response = await openai.responses.create({
  model: "gpt-5.4",
  input: "Audit my product catalog. Return top-10 PDPs to fix.",
  tools: [
    {
      type: "mcp",
      server_url: "https://mcp.boostecom.app/shopify-admin",
      authentication: { type: "bearer", token: process.env.MCP_TOKEN },
    },
  ],
})

The Responses API handles the entire tool loop server-side. You ship the MCP server URL once; ChatGPT discovers tools, calls them, returns the final answer.

→ The hosted MCP endpoint is part of the BoostEcom platform — /marketplace/mcp/shopify-admin ships both a self-host package AND a managed URL.

Why MCP wins long-term

• Provider-agnostic — your MCP server keeps working when GPT-6 ships, when Claude Opus 5 ships, when Gemini 3 ships.
• Tool reuse — the same update_metafield tool used by ChatGPT is used by Claude in the BoostEcom workflow runner.
• Audit log — MCP lifecycle events (tool discovery, invocation, result, error) are structured. Easy to ship to Datafast or your own log drain.

For consumer-facing or team-internal use cases, a Custom GPT is the fastest distribution channel. Anyone with ChatGPT Plus / Team / Enterprise can find your GPT in the picker and chat with your Shopify store.

Setup in 4 steps

1. Open ChatGPT → "Explore GPTs" → "Create a GPT".
2. Configure name, description, behaviour ("you audit Shopify stores").
3. Actions → import an OpenAPI 3.1 spec describing your Shopify tools (a slim wrapper around the BoostEcom API or your own).
4. Authentication: API key or OAuth.

When to pick this path

• ✅ You want to distribute your AI workflow to non-technical users.
• ✅ Your tools are read-only or read-mostly (write actions go through human approval).
• ❌ You need multi-store routing.
• ❌ You need to embed the chat in your own UI.
• ❌ You need conversation history persisted server-side.

For anything beyond a single-store demo, Path A or B beats Custom GPTs.

Same principles as the Claude article: start read-only, observe, gate writes behind human approval. Two ChatGPT-specific notes:

1. Function calling can hallucinate args. Always validate args against your JSON schema BEFORE calling Shopify. ChatGPT may emit first: "ten" when you expected first: 10.

2. Custom GPTs share session memory across users on Team/Enterprise plans when configured that way. Read OpenAI's data-handling docs before distributing a GPT that touches customer PII.

| Dimension | Function calling | MCP via Responses API | |---|---|---| | Setup time | 30 min | 10 min | | Glue code | ~100 lines | ~10 lines | | Streaming | Yes | Yes | | Multi-LLM portability | Locked to OpenAI | Works with Claude, Gemini | | Server requirement | None | An MCP server | | Ideal for | Quick scripts, single use case | Production, multi-LLM, fleet |

→ For BoostEcom we ship MCP first because it's the long-term right answer. Function calling stays available for short scripts and BYOK power users.

The hosted MCP endpoint is the visible part. The hidden parts:

• AI Gateway routing — the platform routes Claude for deep audits, ChatGPT for fast operator chat. Switching is a toggle, not a refactor.
• OIDC + BYOK — your OpenAI key lives where your security policy wants it (BoostEcom's account, your account, or rotated per-org).
• Per-org credit ledger — usage tracked per organization, not per user. Stripe webhook reconciliation built in.
• AI Elements UI — drop-in chat surface with streaming, code blocks, citations. No reinvented Markdown rendering.
• Workflow orchestration — turn one ChatGPT call into a multi-step durable workflow with retries, timeouts, and crash safety.

→ See /features/ai-copilot for the full surface.

Can I use Custom GPTs without writing OpenAPI?

Technically yes — ChatGPT can introspect a free-form tool description. Practically, you'll want OpenAPI for any tool that mutates Shopify state, because the spec doubles as your contract documentation.

What's the cheapest way to test?

Path B (MCP via Responses API) on the free OpenAI tier. The BoostEcom free plan includes daily ChatGPT credits — connect one Shopify store and start chatting in five minutes.

Does ChatGPT's function calling support streaming?

Yes — pass stream: true to chat.completions.create. Tool calls stream as deltas, you assemble them client-side. The Responses API streams natively.

Can I migrate from function calling to MCP later?

Yes. Tool definitions translate cleanly. The orchestration loop collapses from ~100 lines to ~10. We've migrated several DIY clients onto the BoostEcom MCP layer in under an hour each.

What about ChatGPT Plus's "Memory" feature?

It's separate from your integration's session state. Memory is shared across the user's ChatGPT sessions. Your Shopify integration's conversation context lives in your backend (or in BoostEcom's session store).