Seven concrete tasks an AI assistant handles through the Shopware MCP server, with example prompts, tools, permissions and common mistakes.
With the Shopware MCP server, an AI assistant such as Claude or ChatGPT handles real store tasks: building reports, checking stock, finding customers, booking shipments, filling in product copy, creating products and changing settings. The server has been a fixed part of the Shopware core since version 6.7.14 and works with the permissions of an integration you create in the Administration.
Every change runs as a preview first. Shopware calls this a dry run: the model shows what it would change and only writes after a second, confirmed call. The one exception is the image upload, which saves immediately.
The value depends less on the AI than on clean roles. If you create a separate integration with a tight allowlist for each task, you can use read-only tasks on day one and grant write access step by step.
This article is for store owners, e-commerce managers and customer service leads who run a Shopware 6.7 store and want to know what an AI assistant can actually take over today. The basics of the protocol, clients and setup are covered in our article MCP Server for Shopware. Here we focus on practice: seven tasks, each with an example prompt, the tools running in the background, the permissions required and the points where you should be careful.
The MCP server is an interface inside your store that lets an AI assistant read data and, after approval, change it. It runs at /api/_mcp and uses the Admin API, the management interface of Shopware. Shopware introduced it in version 6.7.11 behind a feature flag, and since 6.7.14 it is always on. According to the release notes, version 6.7.16 from November 2026 improves, among other things, how large tool responses are handled. Until Shopware 6.8 the server is labeled experimental, so tool names may still change.
Shopware does not ship special commands like “daily report” or “find customer”. Instead it provides generic tools that work with any entity, meaning any data type such as product, order or customer:
shopware-entity-search searches records, shopware-entity-read reads a single record, shopware-entity-schema shows the fields of an entity.shopware-entity-aggregate calculates sums, averages, counts and trends directly in Shopware without transferring rows.shopware-entity-upsert creates or updates, shopware-entity-delete deletes, both as a preview by default.shopware-order-state changes the state of an order, payment and delivery, shopware-media-upload uploads images from a URL, shopware-system-config-read and shopware-system-config-write read and change settings, shopware-theme-config handles the theme.On top of that come resources with reference data such as shopware://sales-channels, shopware://currencies or shopware://state-machines. From these the model learns, for example, which state transitions are allowed for an order.
MCP is built for dialogue. Shopware limits calls on the admin endpoint to 300 per minute and 1,000 per ten minutes. For importing 20,000 prices, the Sync API via the Shopware Admin API remains the right tool. Processes that run overnight without people also belong in an interface or a workflow, not in a chat.
The most common task is also the safest: asking for numbers. In many stores today this means exporting orders, filtering in Excel and building a pivot table. With MCP you ask the question directly.
Example prompt: “How much revenue did we make yesterday per sales channel, how many orders were there, and what was the average order value? Compare with the same weekday last week.”
The model reads the sales channels from the resource, calls shopware-entity-aggregate on the order entity and builds the sum and average of amountTotal plus a count, filtered by date. The weekly comparison adds a second call. Shopware returns only the results, not individual orders. That keeps the answer small and protects customer data.
The integration only needs read access to orders and sales channels. In the allowlist, shopware-entity-schema, shopware-entity-search and shopware-entity-aggregate are enough. Write tools are left out entirely.
If you hold your own inventory, you lose revenue through empty shelves and capital through full ones. An AI assistant does not replace an inventory system, but it can deliver a list in seconds that someone would otherwise have to filter and sort.
Example prompt: “Show all active products with stock below 10, sorted by quantity sold in the last 30 days. For each product, estimate how many days the stock will last.”
The model uses shopware-entity-search to find products with the filters active and stock below 10. It determines the quantity sold through an aggregation on order line items. The model calculates the coverage itself: stock divided by average daily sales. You should understand this calculation, because it does not come from Shopware.
Read access to products and orders is enough. If the model should also correct stock after a delivery, shopware-entity-upsert with update rights on products is added. In practice we split this into a separate integration for the warehouse.
In customer service, the answer is rarely what takes time. Gathering the information does: finding the customer account, opening the order, checking the payment status, copying the tracking number. That is exactly what the assistant takes over.
Example prompt: “Find the customer with order number 10482. Summarize her last three orders with date, line items, payment status, delivery status and tracking number.”
The model searches for the order with shopware-entity-search, reads it with shopware-entity-read including line items, payments and deliveries, and then looks for other orders from the same customer. From the result it writes a short summary your team uses to reply.
Read access to order, customer and delivery, plus the three read tools. Shopware explicitly recommends a read-only integration for exactly this case. The reason is prompt injection: order notes, customer names and address lines come from customers and can contain text meant to push the model toward other actions. Permissions limit the damage, but they do not prevent the deception.
From here on the assistant changes data. This task is still a good early step, because the shopware-order-state tool only knows state transitions that Shopware allows anyway.
Example prompt: “Set the delivery of orders 10501 to 10512 to shipped. Show me the list with the current status first.”
The model reads the allowed transitions from shopware://state-machines and calls shopware-order-state with deliveryAction: ship, first with dryRun=true. You see which orders would change and how. Only after your confirmation does the same call follow with dryRun=false. One call can change order, payment and delivery status together, for example for a cancellation with a refund.
The role needs order:read and, depending on the action, update rights on order, payment or delivery. Someone who may only book shipments gets only order_delivery:update. A cancellation then fails at the permission level, not just at someone’s attention span.
Almost every catalog has missing descriptions in the second language, empty or duplicate meta titles and incomplete properties. A language model fills such gaps well when it gets clear instructions.
Example prompt: “Find all active products in the Garden Furniture category without a meta title. For the first 20, suggest a meta title of no more than 60 characters, pattern: product name, key feature, brand. Show the suggestions as a table and do not save anything yet.”
The model searches for the products, reads name, properties and description and drafts suggestions. After your review it calls shopware-entity-upsert, first as a preview with the old and new values, then confirmed. For translations you name the target language, and the model writes into the translation for that language.
Read and update rights on products, no rights on orders or customers. The allowlist contains the read tools and shopware-entity-upsert, but not shopware-entity-delete.
New goods often arrive as a spreadsheet from the supplier: SKU, name, purchase price, recommended retail price, image URL. Creating each product in the Administration takes a few minutes. For a handful of products an import interface is not worth it, but for 15 products a week the assistant is.
Example prompt: “Create these eight items as inactive products: name, product number, stock 0, manufacturer Gardenworks, standard tax rate, gross price from the sheet. Show me the preview first. Then upload the image from the Image URL column as the cover image.”
Shopware works with IDs internally. The model therefore first looks up the ID of the tax rate and the manufacturer, reads the currency from shopware://currencies and then builds the shopware-entity-upsert call. The preview shows every product with all fields. After approval, shopware-media-upload uploads the images and sets them as cover images.
Create and update products, read tax rates, manufacturers and currencies, plus media:create for the upload. Because the upload saves without a preview, only this integration gets the media tool.
Shopware’s system settings are extensive and spread out. Anyone who wants to know how many products a category page shows or which sort order is the default clicks through several menus. The assistant reads this in one step.
Example prompt: “Which settings apply to product listings in the Storefront US sales channel? Change the default sorting to price ascending, but show me what will change first.”
shopware-system-config-read reads all values under the key core.listing, optionally for a specific sales channel. For the change, shopware-system-config-write shows a before and after comparison in the dry run. After confirmation it saves the new value. For logo and colors there is shopware-theme-config, also with a preview.
system_config:read to read, system_config:update to change. These rights belong in an integration for a person who already maintains settings today, usually the store manager or the agency supporting the store.
The most important decision comes before the first prompt: which integration may do what? Shopware recommends one integration per task or per role in the team. The table shows how we split the seven tasks in projects.
| Task | Integration | Tools in the allowlist | Writes data |
|---|---|---|---|
| 1 Daily report | Reporting | schema, search, aggregate | no |
| 2 Stock | Reporting or warehouse | schema, search, aggregate, upsert for warehouse | warehouse only |
| 3 Customer service | Service read-only | schema, search, read | no |
| 4 Shipping and cancellation | Service status | search, read, order-state | yes, with preview |
| 5 Product copy | Catalog | schema, search, read, upsert | yes, with preview |
| 6 New products | Catalog | schema, search, read, upsert, media-upload | yes, images without preview |
| 7 Settings | Store management | system-config-read, system-config-write | yes, with preview |
The three discovery tools, which a client uses to find more tools and enable toolsets, are added to every allowlist. Important: enabling a toolset grants no permissions. The integration’s allowlist remains the boundary. An MCP access should never be tied to an admin account, because administrators bypass the user allowlist.
A hypothetical example based on our experience: a retailer with 1,500 orders a month creates a weekly report every Monday. Today that takes 60 minutes, with the assistant 10. The service team handles 400 inquiries a month and saves about 2 minutes per inquiry on gathering information. Together that is a little over 3 hours for reports and about 13 hours in service, so roughly 17 hours a month.
On the other side are the setup with roles, allowlists, prompt templates and training, which in our experience costs 1,500 to 5,000 euros, and the running cost of the AI clients, around 20 US dollars per user per month for Claude Pro or ChatGPT Plus. At an internal rate of 40 euros per hour, a setup at the lower end pays for itself after two to three months. For a small store with 100 orders a month, mainly the read-only part pays off, and it needs hardly any setup.
The seven tasks work, but not without supervision. These are the limits we see most often in practice.
Most Shopware partners are based in Europe, including us. For a US store that means planning around the time difference: between Bremen and the East Coast there are six hours, so reviews of previews and logs work best in the US morning. Sales tax is the second point. Shopware can store tax rules per state, but keeping local rates current by hand is not practical, so most US stores connect a tax service through an integration. Product creation via MCP should leave tax calculation to that service instead of letting the model pick rates. Customer data sent to Claude or ChatGPT falls under the privacy laws of the states you sell to, such as the CCPA in California, and under the provider’s business terms. Prices in this article stay in euros because they come from our European projects.
For retailers who do not want to maintain their own roles and prompts, there are alternatives. The Shopware Copilot in the Administration uses the same MCP endpoint with the permissions of the logged-in user and asks before making changes; its agentic features are part of the Intelligence+ add-on subscription. The experimental SwagMcpMerchantTools package brings ready-made tools such as merchant-revenue-report or merchant-customer-lookup, but Shopware reserves the right to drop it in favor of the Copilot. Our article on Shopware AI gives an overview of all AI features Shopware offers.
No. From Shopware 6.7.14, the MCP server is part of the core and always on. You create an integration with a role and an allowlist and connect a client such as Claude or ChatGPT. Plugins like AI Connect add preview views, an audit log and ready-made configurations.
Only if you allow it. All write tools except the image upload run as a preview by default, and clients like Claude and ChatGPT present tool calls for approval. If you turn off both, you give up control.
The daily report. It needs only read access, is useful from day one and quickly shows how accurately the model works with your data.
The native MCP server is missing in 6.6. A separate MCP server that talks to the Admin API is possible. We usually recommend updating to 6.7, because a separate server becomes redundant after the update.
The server itself is free. Ongoing costs come from the AI clients. In our experience, setup with roles, allowlists and training costs 1,500 to 5,000 euros.
Want to know which of the seven tasks brings the most value in your store? We check your Shopware version, propose integrations with suitable roles and allowlists, and set up the first prompt templates together with your team. For previews and the audit log we use our plugin AI Connect.
Get in touch through our contact form. If you would first like to know where your store stands today, our free store analysis will help.
Besnik Bilalli is the founder and CEO of solution25, the Shopware agency from Bremen. Before founding the agency in 2017, he built his own online store and grew it to more than 500 orders a day. Since then he has been helping merchants with Shopware projects, from the first budget to ongoing operations. He’s happy to answer questions about this article personally: get in touch.
Responsible for content and updates: Besnik Bilalli. Last updated: November 2026. Next content review: February 2027.