Talking to your store from Claude or ChatGPT

Most store owners already use Claude or ChatGPT for writing and thinking. Workspace MCP connects that assistant to your Orkie workspace, so the same chat window can answer questions from your real data, build a campaign, set up an inbox, or request custom work from the operator team. This post shows what that looks like in practice: the questions you can ask, the actions you can take, and the approval steps that keep an assistant from doing anything you did not clearly ask for.

What the connection is

Workspace MCP is a door between an AI assistant and one workspace. An admin copies a connector address from settings and shares it with the people who should have access. Each person connects Claude, ChatGPT or both through a normal sign-in screen. Every connection is separate, has its own audit history, and can be removed on its own.

Once connected, the assistant discovers what the workspace can do: which modules are installed, which data is safe to read, and which actions each module offers. It does not get a database login, a server, source code, or any other workspace. It gets the reads and actions each module has chosen to expose.

Questions you can ask

The reads run immediately. Some examples that work on a typical store with the pixel, the ad connectors and the store connected:

• "How many sales did we get today, and where did they come from?"
• "Compare best-seller sales for July 1 to 7 with the previous seven days. Break out black products and the sources that sold them."
• "Which landing pages lose the most visitors before checkout?"
• "How much did we spend on Google Ads each day last week, and what did the pixel see in return?"
• "Which products get views but never reach the cart, and how much profit does each one make after cost?"
• "Which organic YouTube videos are associated with orders, and which products have sales but no video?"
• "How many contacts can we legally email right now?"
• "Show me products missing required feed fields."
• "Is Google Ads working? What is the one thing I should fix?"

The assistant says where an answer came from: saved workspace data, a live request to the provider, or both. It says how fresh the data is and whether a live request used provider quota. When a report is too wide for a single answer, it tells you which pieces are still needed rather than dropping half the result.

You can also ask it to save a report so it can be rerun, and to compare periods with the change per cell.

Actions you can take

Anything that creates, edits, sends, publishes, syncs or uploads is a change, and changes follow a different path from reads.

For a low-risk change, a clear instruction in the chat is enough. "Add an internal note that this customer asked for a manager callback." The assistant shows the exact effect and does it.

For anything with money or a public result, the assistant shows exactly what will happen and asks once. You reply yes, or send it, or another clear confirmation. Some examples:

A few rules hold everywhere. Every campaign or ad the assistant creates starts paused; making it live is a separate step. The assistant can only send traffic to a page your account already advertises or a path on your own site; it can never choose a destination of its own. It cannot upload a logo. It cannot see tokens, secrets or raw provider responses. Vague money instructions, such as "spend whatever you think is right," are refused rather than interpreted.

A second approval for permanent changes

Some changes get an extra lock. A permanent CRM change, for example deleting or merging customer records, is frozen as an exact proposal: the operation, the arguments, the risk and an expiry. A different workspace admin approves it in settings before it runs. The assistant cannot approve its own proposal, and the same person cannot approve both sides.

Recurring actions work the same way. "Set campaign 123 to 500 every Monday" becomes a task with the exact operation saved on it. A person opens the task, reads the saved action, ticks two acknowledgments, and approves the automation. From then on that exact action can repeat on schedule. Change the amount or the campaign and the approval no longer matches; a new one is needed.

Having custom work built

Sometimes the answer to a question is "the workspace does not do that yet." A module might not expose the report you want, or you want a feed rule nobody has written, or a storefront widget for a specific need.

That is where the operator team comes in. From the same chat you can create a Projects task describing the work, assign it to the Orkie team, and track it. The assistant reads the task context, the operators build it, and the result shows up as a new capability in the workspace. Custom code is a request you make, not a thing you have to write. This is the "operators, not consultants" part of the model: the humans do the work rather than advising you on how to do it.

What it cannot do, on purpose

The limits are worth stating plainly, because they are the reason it is safe to give a marketing lead this access:

• It reaches only the workspace that issued the connection.
• No code, shell, files, deployments, database queries or arbitrary websites.
• External systems such as Shopify, Google Ads, YouTube and email work only through installed modules that advertise a safe operation.
• Provider actions labeled add, update, submit or dismiss always require approval, even if a module calls them reads.
• An approved action can still fail if the provider rejects it or the owning module's own safety rules block it.
• Access is revoked in one click and takes effect on the next request.

Where to start

Connect, then start with a read: "Tell me what data and actions are available in this workspace." Ask for a report on last week. Then try one small change with a preview. Most owners settle into a rhythm within a few days: questions in the morning, a few bounded actions in the afternoon, and a task filed for anything that needs a person.

Want to ask your store a question? Talk to us.

---

Want this for your business? Talk to us How the pixel works