MCP for Agencies: Running AI Across Multiple Client Accounts

Why one login is not enough for an agency

Vendor integrations are built for the person who owns the account. Shopify's official integrations for Claude, ChatGPT, Perplexity and Muse connect one store at a time, even when the stores sit in the same organization. To switch, you ask the AI tool to connect a different store. It disconnects the current one and starts a new authorization, and Shopify advises opening a new conversation afterwards.

For a merchant with one store, that is fine, and Shopify charges no extra fee for the connector. Across twelve clients it means re-authorizing all day.

An agency needs things the single-account model does not give it: a wall between clients, teammates who can read without being able to write, and a way to cut off a contractor without phoning every client.

How a multi-client MCP setup works in PopMCP

The unit is a workspace. One client, one workspace, with that client's connections and usage data inside it.

Provider credentials stay in PopMCP, encrypted at rest with AES-256-GCM. They are never passed to Claude or ChatGPT. Every AI client uses the same URL, https://app.popmcp.com/mcp, signs in with OAuth, and reaches only the connectors the person ticked on the consent screen.

For the mechanics of several accounts behind one endpoint, see multiple accounts, one MCP.

Setup playbook

1. Count before you pick a plan. One connection is one authorized provider account. Twelve clients with four tools each is 48 connections and 12 workspaces.
2. Create a workspace per client. Name it the way your team talks, for example "Acme AU". list_connections shows each connector with its workspace, so a clear name helps the model pick the right one.
3. Connect each client's accounts inside that workspace. Use access the client has granted to your agency. Do not borrow a founder's personal login.
4. Decide who writes. On Starter and Studio there is nothing to set: the owner has full access and every other teammate is read-only. On Scale, set read-only or full access per teammate and connection, and limit each teammate to their own clients' workspaces.
5. Have each person connect their AI client. Paste the URL, sign in, tick the connectors. The Claude setup shows the steps. Each AI client holds its own grant, so connecting Claude does not connect ChatGPT.
6. Name the client in every prompt. With several Shopify connections authorized, the model has to choose one. Say which. Try a read first, because on a full-access connection a write runs as soon as the model calls it.
7. Look at usage. Usage analytics show which tools ran, on which connection, for which teammate.

Prompts that work across a retail client's four tools look like this:

More examples are in MCP use cases for ecommerce.

PopMCP does not schedule anything. A Monday report happens because someone asks for it on Monday. For jobs that must run on a timer, use a workflow tool such as n8n.

Offboarding a contractor

Remove the member. It applies on their next request. On Scale you can narrow which workspaces they reach instead of removing them.

Their AI client only ever held a PopMCP sign-in. It never held the client's Shopify or Klaviyo credential, so nothing of the client's is sitting in a config file on the contractor's laptop.

Two limits. There is no screen that lists or removes one AI client's grant on its own. The controls are removing the member, disconnecting the connector, or running consent again with fewer connectors ticked. And if the contractor had their own staff login to a client's store or ad account, that is outside PopMCP. Remove it with the client.

Pricing example: a 12-client agency

Take 12 retail clients, each connecting Shopify, Klaviyo, Meta Ads and GA4. That is 12 workspaces and 48 connections, with a team of 6. Prices are USD per month, from the pricing page on 5 October 2026.

Extra connections are $4 a month and extra workspaces are $12 a month. Annual billing charges 10 months for 12, add-ons included: $1,130 a year on Studio, or $1,990 on Scale.

Studio is the cheaper fit if the owner does the write work and everyone else reads and reports, since Studio teammates are read-only. Scale is the fit if several people need to write, or if contractors must be limited to certain clients. Scale also adds the IP allowlist and a year of usage history.

Tool calls are never metered on any plan, so a busy month does not change the bill.

What PopMCP does not do for agencies

• No schedules or triggers. It answers when an AI client calls. It does not run a report at 9am or watch for a low-stock event.
• No approval step. PopMCP does not hold a write for a person to approve. Any pause comes from the AI client. Claude, for example, shows a tool approval request unless you set that tool to "Allow always".
• The owner cannot be read-only. Read-only applies to teammates.
• One connection per provider per workspace. A client with three Shopify stores needs three workspaces.
• A focused catalog. 100+ integrations across commerce, ads, analytics, finance and CRM, which is far smaller than Zapier's. You can request an integration.
• No SOC 2 report. Sign-in is Google or a one-time email code, and SSO is on request for Enterprise.

If you manage one store, Shopify's own connector is the simpler choice. If a client's security team needs more than this list offers, how to choose an MCP provider has the questions to put to any vendor, us included.