Vibe coding in Russia: AI models and tools that work without a VPN or a foreign card
Vibe coding is when you describe a task in plain words and an AI agent writes, runs and fixes the code. It takes two things: an agent (Cline, Kilo Code, Claude Code, OpenCode, Codex) and a model that writes the code. The agents themselves are free or distributed as open source, but Claude and GPT models cannot be paid for directly from Russia. With RuAPI the agent connects to Claude, GPT, DeepSeek and Gemini with one key: no VPN, no subscription, paying in USDT only for the tokens you use.
Updated:
API Provider: Anthropic
Anthropic API Key: sk-...
Use custom base URL: https://www.ruapi.ai
Model: claude-sonnet-5-5
These are the fields in the settings of Cline, a free agent for VS Code. Connect Claude through the Anthropic provider and other models through OpenAI Compatible; replace sk-… with your RuAPI key. Kilo Code, OpenCode, Claude Code and Codex are set up in a similar way — pages with instructions are listed under “See also”.
What vibe coding is, and what gets in the way of doing it from Russia
Vibe coding is development through a dialogue with an agent.
The term vibe coding was coined by Andrej Karpathy in a post on X on 2 Feb 2025: that is what he called a way of programming where you “give in to the vibes” and barely look at the code. In practice you write what the result should be, the agent creates files, runs the project and reads the errors, and you check the result and refine the task. The tools fall into two groups: agents that work in your project (Cline, Kilo Code, Claude Code, OpenCode, Codex, Cursor) and cloud app builders (Lovable, Bolt.new, v0, Replit).
The popular tools are tied to a foreign subscription.
Since 4 Sep 2026 Cursor has stopped responding to users in Russia (according to RBC). Google Antigravity does not work in Russia: the country is not on the supported list (antigravity.google/docs/faq, 11 Nov 2026). Claude and ChatGPT subscriptions, like the cloud app builders, are as a rule paid with a foreign bank card, and Russian bank cards do not go through abroad.
The agent is free — what you pay for is the model.
Cline, Kilo Code and OpenCode are open-source programs: they charge nothing and send requests to whatever API address you give them. What remains is getting a key to a strong model. Anthropic and OpenAI do not serve Russia, so you need a gateway: RuAPI opens from Russia without a VPN, and the balance is topped up in USDT, which is bought for rubles via SBP.
How to start vibe coding: 4 steps
The example uses Cline in VS Code: it is the easiest start if you have not worked in a terminal before.
Sign up for RuAPI
Create an account at www.ruapi.ai with email, Google, Yandex or Telegram. A new account usually gets $0.05 right away: enough to try the agent on deepseek-flash, claude-haiku-5-5 or gpt-6-luna.
The other models unlock after your first top-up — from 5 USDT (TRC20, BEP20, TON, Polygon) or USDC. How to buy USDT for rubles via SBP is shown at docs.ruapi.ai/en/buy-usdt. Foreign bank cards are accepted too.
Create a key with a spending limit
Console → “API keys” → “Create key”. An agent makes many requests in a row, so give the key a budget: turn off “Unlimited” and enter an amount — the key will not spend more than that, even if the agent gets stuck in a loop.
Install an agent and connect a model
- Install VS Code and the Cline extension from the VS Code Marketplace.
- For Claude: API Provider — Anthropic, key sk-…, turn on Use custom base URL and enter https://www.ruapi.ai (no /v1), model — claude-sonnet-5-5.
- For DeepSeek, GPT, Gemini and the rest: API Provider — OpenAI Compatible, Base URL https://www.ruapi.ai/v1, the same key, Model ID — the exact name from the pricing page, e.g. deepseek-flash.
- If you prefer the terminal, take Claude Code, OpenCode or Codex; Kilo Code works in JetBrains.
Describe the task and check the result
- Start small: “Make a to-do list page in HTML and JavaScript, store the data in localStorage”. First ask for a plan (Plan mode in Cline), then allow it to execute (Act).
- Run what you got and give the error text back to the agent — it will fix it itself.
- Keep the project in git and commit after every successful step: if the agent breaks something, the changes are easy to roll back.
- Do not give the agent passwords or production keys, and look through the commands it asks you to approve.
Cline setup: docs.ruapi.ai/en/integrations/cline · addresses and keys: docs.ruapi.ai/en/quickstart · caching: docs.ruapi.ai/en/prompt-caching
AI and tools for vibe coding
Agents that take your own key
Cline — VS Code
Kilo Code — VS Code, JetBrains, terminal
Claude Code, OpenCode, Codex — terminal
Zed — an editor with a built-in agent
Models
claude-sonnet-5-5 — hard tasks
gpt-6.1-sol — an alternative for hard tasks
deepseek-flash — cheap routine work
gemini-4-argon — long context
Payment
only for tokens, no subscription
USDT / USDC from 5 USDT
a budget for each key
the balance does not expire
What AI models for vibe coding cost
The models people most often connect to agents. Prices are in US dollars per 1M tokens, from the RuAPI pricing page. You are charged exactly for the tokens the agent used.
| model | input $/M | output $/M |
|---|---|---|
| deepseek-flash | $0.17 cache read $0.0120 | $1.44 |
| gpt-6-luna | $0.12 cache read $0.0120 >272K: $0.24 | $0.60 >272K: $0.90 |
| claude-haiku-5-5 | $0.12 cache read $0.0120 >100K: $0.60 | $0.60 >100K: $3.00 |
| claude-sonnet-5-5 | $2.40 cache read $0.12 | $12.00 |
| gpt-6.1-sol | $2.40 cache read $0.12 >272K: $4.80 | $12.00 >272K: $18.00 |
| gemini-4-argon | $2.40 cache read $0.12 | $12.00 |
Take one agent step: 30,000 input tokens, 24,000 of them from cache, and 1,000 output tokens. By our estimate it costs about $0.033 on claude-sonnet-5-5 (the 6,000 new tokens are written to the cache at the cache-write price) and about $0.0027 on deepseek-flash. A small app is tens or hundreds of such steps. The session history goes to the model on every step, so start a new conversation for a new task: it is cheaper.
The grey “>272K” row is the vendor’s long-prompt rate: if the prompt is over the threshold, the whole request is billed at that rate.