Doka launches a local AI agent for Russia that needs no VPN, subscription, or cloud

A Russian developer has unveiled Doka, a local AI agent built to remove the biggest barrier for users in Russia: dependence on VPNs, foreign subscriptions, and cloud infrastructure. The app works as a desktop assistant for web research and everyday tasks, and its base model runs directly on the device.

From personal tool to product

Doka started with simple frustration. Over the past year and a half, tools like Cursor, Claude Code, and Copilot have noticeably sped up developers' work, but in Russia they often run into unstable VPNs, complicated payments, and the risk of sudden restrictions. The project's author set out to understand how agentic systems work under the hood. He studied the agentic loop, tool calling, and planners, then built a first working prototype for himself.

At first it was a strictly internal tool with no interface and no onboarding. The agent could search the web, open pages, read their contents, and show its work step by step, from the search query to the finished answer. Only the author could use it: without knowing its internal logic, a newcomer would find the tool almost impossible to navigate.

The turning point came after a live demo for an acquaintance. It became clear that the value lay not in the engineering experiment but in a convenient wrapper for ordinary users.

"What even is this? Can I download it?"

What Doka can do

The current version of Doka is a deliberately narrow product. The app focuses on web search and working with web pages: the user describes a task, and the agent picks its own queries, reads documentation or articles, and returns a structured result.

In essence, it is an agentic interface for everyday web research. The author uses it for preliminary research before tasks, for morning monitoring of ML and backend news, and for working through long, poorly structured documentation.

• Searching and reading web pages without manually switching between tabs
• Quick breakdowns of unfamiliar technologies, libraries, and architectural approaches
• Daily digests on chosen topics such as ML and backend
• Help with long documentation that is easier to hand off to an agent
• Routine research tasks that aren't worth an hour of manual work

The key bet is on a local model. By default, Doka installs Qwen3, and according to the author, that is enough for most real-world scenarios involving searching, reading, and explaining page content.

This is not an attempt to compete with top cloud models on raw quality, but a pragmatic choice in favor of accessibility. Web tasks still require access to websites, but the model itself runs locally, does not send user queries to the cloud, and does not require a monthly fee.

Where the project goes next

The author describes the current version as a minimal starting configuration. Its purpose is not to maximize the feature list but to test demand and usage scenarios. Doka will therefore be developed in stages, moving from the most practical capabilities to more complex ones.

The priority is not abstract "universality" but a set of features that save time here and now. This approach should help the project avoid spreading itself thin on decorative features and figure out faster what will actually keep users coming back.

The roadmap already includes working with local files, analyzing a screenshot or the current screen, long-term memory across sessions, and background tasks that users can launch and leave running alongside their main work.

The app is already available for free and supports Windows 10/11 and macOS 12+. On first launch, the model downloads automatically based on the user's hardware. One important detail: users don't need to open a terminal or set up an environment manually.

What it means

Doka shows how quickly the market for AI tools is changing: the focus is shifting from general-purpose cloud services to local agents tailored to users' specific constraints and habits.

For the Russian market, this is an especially important signal. There is demand not only for the "smartest" model but also for a product that works reliably without a VPN, a foreign bank card, or privacy compromises. If the trend holds, local assistants will become a product category of their own rather than a niche experiment.