The AI ecommerce platform where you confirm every change

Most AI for ecommerce is a text box that drafts a product description. Useful, and small. Here the assistant works the shop itself: it reads your catalogue, your orders and your settings, and it can change them, through the same set of actions the dashboard uses. You say "run a 15% sale on Top Picks until Sunday night" and it does the finding, the checking and the arithmetic.

That only works if you can trust it, so the whole product is built around one rule: the agent proposes, you confirm.

1. 1

   Propose

   You ask in your own words. The assistant reads what is relevant, checks what is already running, and comes back with a proposal rather than a change.

2. 2

   Preview

   Every affected item, the value now and the value it would become, and warnings, such as another rule already discounting the same product. The preview runs read-only: it cannot write.

3. 3

   Confirm

   The confirm button is drawn by the dashboard, never by the agent. The page mints a single-use token bound to the exact change it showed you, runs the preview again before committing, and refuses if the shop moved underneath.

4. 4

   Receipt

   What changed, how many rows, who proposed it and who confirmed it. Agent actions and every money action land in the shop's change log, which you can filter by agent.

5. 5

   Undo

   Changes revert from their receipt in one click, including each step of a multi-step plan on its own. A refund is the deliberate exception: money that has left is final, and the preview says so.

These are the actions it has today, each with its own preview. If an action is not on this list, the assistant tells you so rather than pretending. Translating the shop into another language is one of them; multilingual shops has the detail.

You can give other AI tools access to the shop as named agents, each with its own grants: catalogue, pricing, fulfilment, money, content, design, settings, customers, reports. An agent is not even shown the tools it lacks, and taking a grant away cuts it off mid-conversation.

Agents connect over MCP, the Model Context Protocol, so you can run the shop from Claude itself, on claude.ai or in Claude Code, on both plans. Starting points are one press each:

• Analyst: reads reports and shop state, changes nothing.
• Shopkeeper: orders out of the door, products and stock.
• Merchandiser: prices, sales, coupons and the catalogue.
• Front desk: customers and orders, notes, parcels marked as sent.
• Designer: drafts layout, sections and theme settings; publishing stays yours.
• No template carries permission to move money. That grant is added by hand, by you, and even then every money move is a preview until you confirm it.

Switch it on and a question box sits at the bottom of every page of your shop, with your own shortcuts under it. Shoppers say what they are after, and the assistant finds it, narrows the choice with a tap and fills the bag, over the page they were on. It works with every one of our themes, and your pages and checkout stay exactly as they are.

If your shoppers know what they want and like a grid, leave it off. If they arrive with a question, it answers it.

Some instructions are standing ones: "every Monday, tell me how the week went". You set a routine by asking for it, daily, weekly or monthly, at an hour in your shop's time zone. Every live shop gets a weekly brief: what sold against the week before, orders still waiting, stock running out, and the two or three things most worth doing.

A routine runs unattended, so it never changes anything. What it would do arrives by email as links to a preview you confirm, the same as anything else.

The allowance is counted in what Claude costs us, not in messages, so it means the same thing whichever model a task needs. How the plans compare overall is on the pricing page, and the free ecommerce platform page shows when Pro becomes the cheaper plan.