Telegram bot builds a VkusVill grocery cart from a single sentence

Some projects start with everyday irritation and end up as working products. A Russian developer has published a write-up on Habr about building a Telegram bot that takes a plain text phrase, such as "put together breakfast for two," and builds a VkusVill cart on its own, complete with a ready link for checkout.

The starting point will be familiar to anyone who has ever ordered groceries online. The author describes a typical routine: open the catalog, pick milk from a dozen and a half near-identical cards, then bread, then cheese, then something to go with tea. By the fifth item, any feeling of saved time has evaporated. According to the analytics platform Platforma, Russians spend 19 to 49 minutes a month choosing groceries in mobile apps, and Muscovites spend two and a half times more. Meanwhile, two thirds of shoppers cite saving time as the main reason for moving to online grocery shopping. The paradox is plain: a tool built for convenience becomes a time sink itself.

Technically, the bot is a chain of several components. On the user side it is an ordinary Telegram interface, with no extra apps or registration. The user's message goes to a language model deployed through Yandex Cloud AI Studio. The key technology is function calling, a mechanism in which the model does not just generate text but decides which external functions to call to complete the task. Here, the model "understands" that a breakfast request means a set of specific product categories and calls the VkusVill catalog search functions through its public API. The search results are aggregated into a cart, and the user receives a link where all that remains is to confirm the order.

The author stresses that the project is not affiliated with either VkusVill or Yandex. It is a fully independent development that uses the open interfaces of both services. The code is on GitHub under the Apache 2.0 license, which makes it available for study, forking and commercial use. For developers eyeing function calling and the Model Context Protocol, the project has practical value as a reference implementation: not an abstract example from documentation, but a working product with a real user scenario.

The project matters less in itself than as an illustration of a broader trend. Language models increasingly step outside the chat window and start controlling external services: placing orders, booking tables, searching for tickets. Function calling and MCP turn an LLM from a conversation partner into an operator that can take actions in the real world through APIs. Grocery retail is one of the most obvious candidates for this kind of automation: buying patterns repeat week after week, the assortment is structured, and the user pain of routine catalog scrolling is easy to measure.

The approach has limitations worth keeping in mind, though. Public APIs of large retailers can change structure without warning, so the bot risks breaking at any moment. Personalization is still limited by the model: the bot does not know your dietary restrictions, brand preferences or budget unless you state them explicitly. Finally, there is the question of trust: willingness to delegate grocery choices to an algorithm remains a matter of individual comfort.

Still, the direction is set. If major grocery chains build such interfaces into their official apps, the familiar model of online grocery shopping, endless catalog scrolling with items added by hand, may give way to a conversational format. One independent bot will not, of course, overturn the industry. But it shows clearly that the technology stack for this already exists, is accessible, and does not require a team of fifty engineers. Sometimes one developer tired of searching for buckwheat is enough.