Running Multiple Accounts Through One MCP URL

Why a connector usually means one account

A typical connector signs in to a provider once and stores one grant. If its tools take a date range or an order ID and nothing that names an account, every call runs against the account that signed in.

MCP does not require this. In the spec, a tool's input schema is whatever the server defines, so a server can add an input that selects an account. Whether it does is up to whoever built the server.

The vendors' own docs show the single-account design:

For one store or one CRM this is fine. switch-shop is enough if you move between two stores a few times a week.

Where one account at a time stops working

It stops working for an agency that runs ads, email or stores for a dozen clients, and for a brand with a store per region.

It also fails on any question that compares accounts, such as "Which store had the higher refund rate last month?" With a one-store connector you ask, switch stores, authorize again, open a new chat and ask a second time. Then you compare the two answers yourself.

How PopMCP handles multiple accounts

PopMCP uses three terms, and they are easy to mix up.

The rule to remember is one connection per provider per workspace. A second Shopify store needs a second workspace. Two accounts for the same tool cannot share one.

That makes multiple accounts a paid feature. Free has one workspace. Starter has three, Studio ten and Scale 25, and an extra workspace is $12 a month.

How each account gets connected depends on the provider. For Google Ads, Meta Ads and GA4 you authorize once, then choose which ad accounts or properties to connect and which workspace each belongs in. Shopify goes store by store: each store is authorized on its own and becomes its own connection.

How the model picks the right account

1. The model calls list_connections. For each connection it gets an ID, the provider, the team and workspace names, a label, and whether the connection is read-only.
2. Every provider tool takes a connection input. With one Shopify connection it is optional. With two or more it is required, and a call without it returns an error that lists the options.
3. The call runs against that one connection.

So "Compare last week's orders in Acme AU and Acme US" becomes the same tool called twice, once per connection, with one answer at the end.

Workspace names do real work here. The label the model reads is built from the provider, your team name and the workspace name. A workspace called "Store 2" gives it nothing to match against "the US store".

Who can reach which account

Each AI client signs in on its own and gets its own grant. On the consent screen you tick the connectors that client may reach, so Claude can have every client workspace while one Cursor install has a single store.

For teammates, Scale adds per-member workspace access, and the owner sets each teammate to read-only or full access per connection. On Starter and Studio, teammates other than the owner are read-only.

To take an account away from an AI client, run its sign-in again with that connector unticked, or disconnect the connector in PopMCP. There is no separate screen that lists and cancels client grants.

Provider credentials stay in PopMCP, encrypted at rest with AES-256-GCM. The AI client never holds a store's API secret.

How to set it up

1. Create a workspace for each store, brand or client. Name it the way you would say it in a prompt.
2. Connect each account inside its workspace. For Shopify, that is one authorization per store.
3. Add PopMCP to your AI client once. The URL is https://app.popmcp.com/mcp for every workspace. Here are the steps for Claude and ChatGPT.
4. Tick the connectors that client should reach on the consent screen.
5. Test with a read. "Run list_connections, then give me yesterday's order count for each Shopify store."

What it costs

Connections are pooled across workspaces, and tool calls are not metered.

On Starter, Studio and Scale, an extra connection is $4 a month and an extra workspace is $12. The full table is on the pricing page.

When you do not need this

With one store, or two you switch between now and then, start with Shopify's own connector. Shopify charges no extra fee for it, it installs from Claude's directory, and Shopify supports it. A ChatGPT app that supports more than one connected account may be enough for a personal login and a work login.

A hosted setup earns its cost when the accounts multiply, when you want them in the same conversation, or when different people and AI clients should see different ones. MCP for agencies covers that case.

Prompts that keep accounts straight

Name the account in every request, using the workspace name: "In Acme US, list the ten most recent orders."

Before a change, ask which connection the model will use. PopMCP does not hold a write for approval, and a write runs when the model calls it. The pause has to come from you or from the AI client. ChatGPT asks for confirmation on write actions by default, and Claude can ask before it uses a tool.

For a client where the AI should only report, do not rely on the prompt. On Scale or Enterprise, set the teammate to read-only for that connection. The owner's own access cannot be made read-only, so an owner working alone depends on the client's confirmation step.