AI assistants such as Claude and ChatGPT already fill baskets in shops that have given them official access. We build an MCP server for your shop that gives this access to any assistant supporting the open MCP standard. The assistant searches your catalogue and builds the order with promotions taken into account, while the customer places and pays for the order on your website.
The open MCP standard (Model Context Protocol) describes how an AI assistant connects to an outside system and what it is allowed to do there. The shop publishes an MCP server with a list of operations: find a product, show promotions, add to the basket, pick a delivery time. The customer connects this server in Claude, ChatGPT or another assistant and writes in plain language: “get lactose-free groceries for the week, keep it under 80 euros”.
From there the assistant goes through the catalogue by itself and compares prices per kilo. The customer checks the finished basket and pays for it in your shop, as usual.
The standard is not tied to one platform. Besides Claude and ChatGPT, the server also works with agents that people run themselves: OpenClaw, Cursor, Codex and other MCP clients. The shop does not have to be listed in any directory for this.
Without such a server, the assistant opens the website like a browser and looks for the right buttons on the page. This is slow, and every layout update breaks something. The shop does not even know what the agent sees and what it does on the customer’s behalf.
Large US retailers started at the end of 2025 and now run official apps inside AI assistants. In European grocery retail there are only a handful of such cases so far. In the Baltics, as far as we know as of October 2026, no retailer has yet given AI agents official access.
In ChatGPT since December 2025, in Claude since April 2026.
Its ChatGPT app has been live since late March 2026, but Walmart kept account sign-in and payment on its own side.
Has sold through ChatGPT in France since March 2026 and in Belgium since August. Orders are paid for on the chain’s website.
Since August 2026 the chain has had its own MCP server with about 40 tools, from search and substitutions to coupons and delivery slots. The customer pays via a link to the website.
Every shop on the platform has an MCP server.
Less than a month after the launch, Silpo confirmed that customers were already placing orders through AI agents. What surprised the chain was that some of these customers had never bought from it before. Another Ukrainian shop has since launched its own MCP server. Silpo does not disclose exact sales figures.
In March 2026 OpenAI dropped its original model, in which the customer paid right in the chat. Payment now happens on the shop’s website or in its own app, and the assistant only fills the basket.
In July 2026, people who came to US retailers’ websites from AI sources bought 60% more often than other visitors, and their revenue per visit was 53% higher. These are Adobe figures for all US online retail. There are no such figures for groceries or for Estonia.
AI agents visit shop websites even without your consent. According to HUMAN Security, traffic from AI agents and agentic browsers grew by 7,851% in a year. With an official server, the agent gets current prices and stock from you, and you set the request rate.
In our experience, putting together a weekly basket by hand takes from 40 minutes to 2 hours. With an assistant, the customer spends 3–5 minutes checking and paying. Along the way the assistant spots loyalty card prices and personal coupons that are easy to miss in the app.
Server logs show what people search for through assistants and what they cannot find in your catalogue.
An extra option open only to an official server. Anthropic requires the applicant to own or control the API, and OpenAI does not publish unofficial connectors, so a third-party developer cannot get listed without the shop.
The downside is that part of the data about the customer’s conversation with the assistant stays with the AI platform. Nobody has measured yet whether the new channel takes orders away from your app.
We built an MCP server for shopping at a large online grocery store in Estonia. It is a working prototype on the store’s web storefront, built on our own customer account; we propose making the official version together with the store. The server currently has 42 tools. 17 of them change data in the account, such as the basket or shopping lists.
Filters for promotions, country, brand, “lactose-free” and “organic”; loyalty card prices.
Country of origin and ingredients.
Up to 30 items in one call, saved baskets and repeating a past order.
Alternatives for out-of-stock products, and a “remove unavailable” command.
The basket summary shows how much is left to reach a coupon threshold, the minimum order and free delivery.
Available slots.
Previously bought items, favourites and lists, bestsellers, basket suggestions.
The customer places and pays for the order. The server works in any MCP client, including Claude Desktop and Claude Code.
Time spent on a weekly basket by hand and with an assistant (our estimate, not a measurement). With the assistant, the customer only checks the basket and pays.
On 7 October 2026 we ran a real family’s weekly basket through the server. That was an earlier version with 29 tools, still without substitutions for missing products and without basket suggestions.
While filling the basket, the assistant applied promotions, loyalty card prices and a personal coupon by itself. That coupon is easy to overlook in the app.
We write the assistant’s instructions in the customer’s interest. If a server quietly nudges people towards whatever benefits the shop, they notice quickly and stop trusting the channel.
We are building one server for an online household goods shop and another for a building materials shop.
Laiskus Software OÜ, based in Tallinn, mainly does 1C:Enterprise development. In these projects an AI assistant works with the client’s configuration through our own tools. MCP servers for shops are a new line of work for us, and the grocery server is the first one in it.
The assistant acts on the customer’s behalf, so every change in the account goes through the server’s rules.
Before an irreversible action the server shows a preview and issues a one-time token. Without that token the action does not run.
Set on the server. You can limit the assistant to reading, or require confirmation for every write.
The server only contacts shop addresses that were allowed in advance.
No more than one request per second.
The server does not store them; users sign in themselves.
If the customer changes the basket by hand, the assistant sees it.
The server does not do them; it was not built with that ability.
More than 1,200 automated tests, an independent review of each block and a separate security review.
In the pilot we find out which services your website and app run on and build a server for your catalogue and basket. If you want to be listed in the Claude and ChatGPT directories, we help with the applications. After that we look at the first figures together. Timing and scope differ from shop to shop, and we name them after the review.
If the pilot works for you, we sign a development and support contract.
You can order an MCP server on its own or together with other work.
We set up an online shop with a catalogue, basket and payment. An MCP server for assistants can be connected to it from day one.
More about shop development →If your shop has no API, or the API lacks operations, we will develop one. Through it, your mobile app, accounting system and MCP server can all work with the shop.
More about API development →We translate your website, product pages and customer emails into Estonian, English, Russian and other languages.
Discuss translation →You do not need a public API for this. The server uses the same services your website and app run on and opens only the operations the assistant needs. Silpo built its MCP mostly on APIs it already had. We list any missing operations during the review and agree them with your team. If you need an API for other tasks as well, we will develop a full one.
Once the contract is signed, we publish the code of the official server as open source under the Laiskus name. Until then it is not published, and we show the prototype in a live demo. The shop decides whether to keep its own instance of the server running. With open code, anyone who needs it can run a copy themselves, which takes some of the load off the main server. Other rights are set out in the contract. The shop applies to the directories as the owner of the API. Logs of requests to your server stay with you. What the customer writes to the assistant is visible to the AI platform, and your privacy policy should say so.
Payment goes through your website or app, as it does now. The assistant fills the basket and sends the customer to checkout. Under the OpenAI directory rules, the purchase has to be completed on the shop’s domain anyway. The server has no payment part of its own, so your contracts with the bank and payment provider stay as they are.
A person always checks the basket before paying. The server carries out irreversible actions only after a preview and confirmation.
There are no published measurements on this yet. Instacart, for example, launched its own assistant, Clementine, in September 2026 and kept its catalogue open to Claude and ChatGPT at the same time. In the pilot we measure this too, so that the decision to continue rests on your own data.
We name the price after reviewing your API, because the scope of work depends on it.
Tell us which platform your shop runs on and whether your mobile app has an API. If you like, we can show you the prototype at work.
