Skip to content
← Blog
11 min read

Claude Code from Russia: install, pay and run without a VPN

Whether Claude Code works from Russia, how to install it on Windows, macOS and Linux, which Claude models to connect and what a day of work costs. Pay with a Russian card or SBP from a Kumo balance — no Anthropic subscription, no VPN.

claude codeguidepaying from russiavibe coding

Updated October 21, 2026. Claude Code is Anthropic's coding agent that lives in the terminal: it reads your project, writes and edits code, runs commands and tests. From Russia you can use it without a VPN and without an Anthropic subscription — through the Kumo gateway, paying with a Russian card or SBP. This guide covers installing Claude Code on Windows, macOS and Linux, connecting it to Kumo, choosing models and what a day of agent work really costs.

Agent ranking

#1 of 42

Claude Code with Sonnet 5.5 — Coding Agent Index 68.4

Sonnet 5.5 on Kumo

$1.1 / $5.5

per 1M tokens, input / output

One agent task

$1.43

Sonnet 5.5 at xhigh, about 27 minutes of work

Minimum top-up

$2

by card, SBP or crypto

Does Claude Code work from Russia

Yes, but not directly. Russia is not among the countries where Anthropic sells subscriptions and API access, and the company does not accept Russian cards: you cannot buy Claude Pro or Max or top up an API balance from Russia. Claude Code itself, though, is an ordinary program that only needs a server address and a key.

Kumo is an API gateway that accepts requests in the Anthropic protocol and passes them to Claude models. Claude Code connects to it with two environment variables and works from Russia without a VPN, and you pay from your Kumo balance — with a Russian card, SBP or crypto. There is no subscription, no seats and no monthly minimum: you pay only for the tokens the agent spends.

What Claude Code is and how it works

Claude Code is not autocomplete or a chat next to your editor but an agent on the command line. You start it in a project folder and it finds its own way around the repository: looks at the structure, opens the right files, finds where the logic you want changed lives. Then it acts — edits files, runs commands, builds and tests, reads errors and fixes them until the task is done.

The agent shows every change for approval, so you stay in control. Since Claude Code lives in the terminal, it does not care what you write code in: VS Code, JetBrains, Vim or no editor at all. In the independent Artificial Analysis ranking, Claude Code with Sonnet 5.5 is the strongest coding agent on the market.

Installing on macOS and Linux

The official way is Anthropic's one-line installer. It installs a self-updating Claude Code and does not need Node.js. After installing, open a new terminal and check the version.

curl -fsSL https://claude.ai/install.sh | bash
claude --version

Installing on Windows

On Windows the installer runs in PowerShell. Claude Code uses Git Bash to run commands, so install Git for Windows first. You can also work inside WSL — then use the Linux command.

irm https://claude.ai/install.ps1 | iex
claude --version
npm install -g @anthropic-ai/claude-code
claude --version
The path from install to first run: the installer or npm, two Kumo variables and the claude command in the project folder.

Connecting Claude Code to Kumo

Sign up in the Kumo console, top up and create an API key. The most convenient way to connect is the user settings file ~/.claude/settings.json (on Windows %USERPROFILE%\.claude\settings.json): it applies to every project. The Kumo console has a step-by-step Claude Code setup that builds this file with your key — you can simply download it and put it in place.

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.kumorouter.com",
    "ANTHROPIC_AUTH_TOKEN": "kumo_sk_...",
    "ANTHROPIC_DEFAULT_FABLE_MODEL": "claude-fable-5-1",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-5-5",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-5-5",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-5-5",
    "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY": "1"
  },
  "model": "claude-sonnet-5-5"
}

What matters here. ANTHROPIC_BASE_URL is the bare gateway address without /v1: Claude Code appends the path itself. The key goes into ANTHROPIC_AUTH_TOKEN. The four DEFAULT_*_MODEL variables assign models to the Fable, Opus, Sonnet and Haiku slots of the /model menu, so you can switch between them inside a session; Haiku also takes the background calls. The GATEWAY_MODEL_DISCOVERY line lets Claude Code read the gateway's model list at startup.

If you would rather not touch files, environment variables in your shell profile — ~/.zshrc or ~/.bashrc — are enough:

export ANTHROPIC_BASE_URL="https://api.kumorouter.com"
export ANTHROPIC_AUTH_TOKEN="kumo_sk_..."
export ANTHROPIC_MODEL="claude-sonnet-5-5"

# check: the answer should arrive and the request appear in the Kumo console log
claude -p "reply with OK"

Which models to connect to Claude Code

Claude Code speaks the Anthropic protocol, so through Kumo it runs Claude models — all current generations. Here are the four worth assigning to the slots, with their scores in the Artificial Analysis agent ranking.

ModelSlotKumo priceAgent indexWhen to use
Claude Sonnet 5.5Sonnet$1.1 / $5.568.4The everyday default: best result at a mid price
Claude Opus 5.5Opus$2.2 / $1166.0Hard tasks, architecture, large refactors
Claude Fable 5.1Fable$6 / $3062.2The longest autonomous tasks, when price is secondary
Claude Haiku 5.5Haiku$0.06 / $0.341.4Background calls, simple edits, subagents
Price is USD per 1M tokens, input / output, Kumo price list v33. Index is Claude Code with that model at its best reasoning level.

Mind the reasoning level: Sonnet 5.5 scores 68.4 at max, 62.9 at xhigh and 55.0 at high. You change it with /effort inside a session, for example /effort high. All models and agents are compared on the benchmarks page and in the coding models ranking.

What a day of work costs

Artificial Analysis publishes how many tokens Claude Code spends on one task of its benchmark — large tasks of tens of minutes of agent work. Multiply those tokens by Kumo's prices and you get an honest cost per task. We priced uncached input at the cache-write rate — above the real rate — so the estimate errs high.

Model and levelIndexMinutes per taskTask on KumoTask at Anthropic prices10 tasks on Kumo
Sonnet 5.5, max68.487$6.30$14.19$62.97
Opus 5.5, max66.064$7.05$13.04$70.53
Sonnet 5.5, xhigh62.927$1.43$3.33$14.28
Fable 5.1, max62.235$7.61$12.39$76.10
Sonnet 5.5, high55.012$0.51$1.24$5.09
Sonnet 5.5, medium45.99$0.26$0.62$2.59
Haiku 5.5, xhigh41.426$0.11$0.61$1.13
Tokens per task: Artificial Analysis Coding Agents, October 20, 2026; prices: Kumo price list v33. The Anthropic-price figure is Artificial Analysis's own estimate.

Take a working day of eight such tasks on Sonnet 5.5. At xhigh that is $11.43, at high $4.07. At Anthropic's prices the same xhigh task costs 2.3 times as much. A normal day rarely holds eight benchmark-size tasks: smaller edits, code questions and reviews cost cents, so the real bill is usually lower.

The biggest savings lever is reasoning level and model. Sonnet 5.5 at max spends $6.30 and 87 minutes per task; at xhigh, $1.43 and 27 minutes with an index only 5.5 points lower. Haiku 5.5 handles simple tasks for $0.11.

Paying from Russia

The Kumo balance is topped up with a Russian card, SBP or crypto; the minimum top-up is $2. Charges follow usage: the tokens the agent spent leave the balance at the model's rate, which you can see in advance on the models page. Sonnet 5.5 costs $1.1 / $5.5 per million input and output tokens.

Each key can have a spending limit — an agent stuck in a long loop stops there, not at an empty balance. A separate key per project helps: the console log shows which key, which model and what each request cost.

How to work with Claude Code productively

Start with reconnaissance: ask the agent to describe the project — stack, entry points, how it fits together. Then give it a small concrete task: fix a bug, add form validation, write a test. The more specific the request, the fewer iterations and tokens. Cut a big feature into steps and check after each one.

Keep a CLAUDE.md file in the project root: stack and versions, build and test commands, code conventions, what not to do. Claude Code reads it at the start of every session, so you do not repeat yourself. Ten minutes on a good CLAUDE.md pay off on day one — and save the tokens the agent would otherwise spend exploring.

Frequently asked questions

Does Claude Code work in Russia without a VPN?
Yes, when connected to the Kumo gateway: the agent talks not to Anthropic servers but to api.kumorouter.com. A VPN may only be needed for the installer on claude.ai, and installing from npm replaces it.
Do I need a Claude Pro or Max subscription?
No. With Kumo, Claude Code works with an API key and you pay for actual tokens from your balance. No Anthropic account or foreign card is needed.
How do I pay for Claude Code from Russia?
Top up your Kumo balance by card, SBP or crypto — from $2. Claude Code then spends tokens from that balance at the chosen model's rate.
How do I install Claude Code on Windows?
Install Git for Windows, then run irm https://claude.ai/install.ps1 | iex in PowerShell. If the site does not open, install Node.js 18+ and run npm install -g @anthropic-ai/claude-code. Put the Kumo key in %USERPROFILE%\.claude\settings.json.
Can I connect GPT or Gemini to Claude Code?
Through Kumo, Claude Code runs Claude models: the agent speaks the Anthropic protocol. For GPT use Codex, and Gemini, GPT and every other catalogue model connect to Cursor, Cline and OpenCode — see the coding models ranking.
How much does Claude Code cost per month?
There is no fixed price — you pay for tokens. A benchmark task on Sonnet 5.5 at xhigh costs about $1.43 on Kumo; eight such tasks a day come to about $11.4. Light edits and code questions cost cents.
What if Claude Code answers 401 or 404?
404 almost always means an extra /v1 in ANTHROPIC_BASE_URL. 401 means the key never reached the process or an old ANTHROPIC_API_KEY overrides it. Fix it, open a new terminal and check with claude -p "reply with OK".

Further reading

All agent and editor guides live in the vibe coding section. Which models are stronger at code and why — in the coding models ranking and on the agent benchmarks page. Claude Code compared with Cursor, Cline and OpenCode — in the vibe coding tools review, and Anthropic's latest model is covered in the Claude Opus 5.5 article.

Start building on Kumo today

One base URL, one balance, every model — at an effective rate you can see before you spend a token