Russia Now Has 18 Homegrown AI Coding Agents, but "Russian" Means Three Different Things

On 4 September 2026, Cursor disabled its agent and autocomplete for users in Russia and began canceling paid subscriptions with refunds. Engineers who had been working with the agent in the morning were left with an empty editor by lunchtime. Five weeks later, Habr published a review: the author went through every Russian AI coding agent they could find. There are eighteen, and twelve are free to use. But the headline finding is not the count. In this category, the word "Russian" describes three different designs, and each answers differently the question every security officer asks: does the code leave the country?

Three designs under one label

First: in-house models. The company trains the model itself and runs it itself. Code stays inside the vendor's perimeter, and both quality and pace of development depend on the vendor alone.

Second: open models on the vendor's own servers. The vendor takes published weights, deploys them in a Russian data center, and builds an agent layer on top. The code goes nowhere, but someone else's model sets the quality ceiling, and the product only improves when a new version of the weights is released.

Third: foreign models behind a ruble wrapper. The interface is Russian, the contract is Russian, and payment is in rubles. Under the hood, though, is a model that receives requests outside the country. From the outside, the product looks the same as the first two.

The review is no substitute for your own testing, and the three categories on their own do not say which agent writes better code. They say where your source code is physically processed and whose decisions the product depends on.

What the September day without Cursor showed

The Cursor episode exposed more than a leak risk. It is a question of dependence. If a tool is tied to a foreign provider, a single unilateral move by that provider turns a team's workflow into downtime. This is our conclusion, not a quote from the review, but it follows from the design itself. The third group of agents is vulnerable to the same scenario: a ruble wrapper offers no protection if the model provider cuts off access. The first two groups are better protected against that risk, though they pay for it with limits on quality and pace of development.

The review's numbers

• Total Russian AI coding agents: 18
• Free: 12, or two-thirds
• "Russianness" architectures: 3
• Date Cursor cut off Russia: 4 September 2026
• What Cursor did: disabled the agent and autocomplete, began canceling paid subscriptions with refunds
• The main criterion for a security officer: does the code leave the country?

What to check before choosing

Start with four questions for any vendor. Whose weights power the agent: its own, open ones, or someone else's closed ones? Where does inference physically happen, and can you deploy the solution inside your own perimeter? What does the contract say about storing requests and using your code for further training? Does the product have a backup provider if the primary one disappears?

A separate word on free agents. Twelve of the eighteen cost nothing, which is convenient for personal projects. For a commercial codebase, free raises a question about the monetization model: what exactly are you paying with, if not money? We did not find an answer in the text of the review, so ask the vendor directly.

Last point: do not keep the whole team on a single tool. The September day showed that switching to a fallback at the moment of a shutdown costs more than a pilot run in advance.