How to Use Qwen Code: Installation, Account and Model Selection, and API Settings

Qwen Code is an open-source (Apache-2.0) AI coding agent released by Alibaba’s Qwen team. Its usage consists of four steps: “install → run qwen in the project → configure the connection target with /auth → request a task.” However, the free Qwen OAuth tier ended on April 15, 2026, so logging in according to instructions published before then no longer provides free operation. One more point Korean users should know first: Qwen Code has no Korean UI. This page explains, in order, how to get started based on the current /auth menu, configure Korean responses, use Claude and GPT with a custom setup, and use frequently needed commands. The latest version is v0.24.7, published to npm on September 29, 2026, and releases appear at least weekly, so check the installed version with qwen --version.

Installation

These are the commands from the official README. The standalone installation script means you do not need to prepare Node.js yourself; Node.js 22 or later is required only when installing with npm.

# macOS / Linux (공식 단독 설치 스크립트)
curl -fsSL https://qwen-code-assets.oss-cn-hangzhou.aliyuncs.com/installation/install-qwen-standalone.sh | bash

# Windows (PowerShell)
irm https://qwen-code-assets.oss-cn-hangzhou.aliyuncs.com/installation/install-qwen-standalone.ps1 | iex

# npm (Node.js 22 이상 필요)
npm install -g @qwen-code/qwen-code@latest

# Homebrew (macOS / Linux)
brew install qwen-code

Reopen the terminal after installation so environment variables take effect. Qwen Code was originally based on Google Gemini CLI v0.8.2, but since v0.1 it has been developed independently rather than synchronized with the original. Gemini CLI settings and free quotas do not apply unchanged.

First run and Korean responses

cd /path/to/your-project
qwen

# 세션 안에서:
/language output Korean   # 모델 답변을 한국어로 (UI는 한국어 미지원)
/auth                     # 공급자와 API 키 설정

The built-in UI languages are limited to Simplified Chinese, English, Russian, German, Japanese, Brazilian Portuguese, French, and Catalan, and Korean is absent from the repository’s UI translation files (checked October 1, 2026). Menus therefore appear in English, and this page quotes the English wording shown on screen. The model response language is separate from the UI, so it can be fixed to Korean with /language output Korean.

In addition to the terminal, there are a desktop app, a browser-based web UI (qwen serve --open, experimental), VS Code, Zed, and JetBrains integrations, and headless execution for scripts and CI (qwen -p "..."). They are all included in the same free repository, but inference costs are separate.

Choose a connection target with /auth

Opening /auth (alias /login) displays the “Select authentication method:” screen, whose top-level choices are three options. If you try to choose Qwen OAuth, an “Discontinued — switch to Coding Plan or API Key” notice appears.

The three ModelStudio options are not three ways to pay the same bill. The documentation uses a separate endpoint and separate key for each, and they do not work if the key type and baseUrl do not match. As of October 1, 2026, Coding Plan was marked “limited quantity, first come first served; replenished daily at 00:00 (UTC+8),” so sign-up may be unavailable depending on the day. A task in Coding Plan is counted as multiple model calls; Alibaba says 5–10 calls for a simple task and 10–30 or more for a complex task are typical. The English page Qwen Code pricing covers the pricing comparison in detail.

Using Claude or GPT: Custom Provider in settings.json

The Qwen Code authentication documentation recommends ~/.qwen/settings.json’s modelProviders for third-party connections such as OpenAI, Anthropic, Google, OpenRouter, and custom endpoints. For Kunavo, use the following.

{
  "modelProviders": {
    "openai": [
      {
        "id": "claude-sonnet-5",
        "name": "Claude Sonnet 5 (Kunavo)",
        "baseUrl": "https://api.kunavo.com/v1",
        "description": "Kunavo, OpenAI 호환",
        "envKey": "KUNAVO_API_KEY"
      }
    ]
  },
  "security": {
    "auth": {
      "selectedType": "openai"
    }
  },
  "model": {
    "name": "claude-sonnet-5"
  }
}

# 키는 settings.json에 직접 쓰지 말고 .qwen/.env 나 환경 변수로
echo 'KUNAVO_API_KEY=sk-kn-...' >> ~/.qwen/.env

Four things that save debugging time:

• baseUrl only goes through /v1. The model provider reference documentation explicitly says to “specify the API’s /v1 root, not the full /v1/chat/completions path; the SDK appends the request path.” Including the full path causes a 404, not an authentication error.
• Where to put the key. Qwen Code reads the key from the environment variable whose name is specified in envKey. The priority order is shell export, .env files (.qwen/.env recommended; only the first file found), and settings.json’s env; the last option stores the key in plain text and is not recommended.
• The configuration file takes precedence over CLI flags. The order is /auth input → selected modelProviders entry → CLI arguments → environment variables → settings.json. This is why --openai-base-url may appear to be ignored. The security.auth.apiKey and security.auth.baseUrl mentioned in older explanations are scheduled for deprecation.
• Start with Chat Completions. If you omit wireApi, the Chat Completions format is used. Switching to "responses" disables both automatic endpoint detection and automatic failover. Changes to modelProviders take effect immediately even in a running session (they appear after reopening /model).

The Kunavo catalog has no Qwen text models. This is not a way to use Qwen cheaply; it is a way to use Claude and GPT inside Qwen Code with one prepaid balance. Kunavo has also not verified running Qwen Code against its own endpoint. As with the English Qwen Code configuration page, this configuration was written by consulting the official documentation. Keep your current route in place while testing.

First task and frequently used commands

The first request shown in the README is something like “Explain this repository and tell me where to start.” It verifies file reading and tool calls at once. A greeting alone cannot reveal connection problems.

An automatic approval mode such as yolo is explicitly warned by the documentation to be used only in an environment you trust, a sandbox, or one you can afford to discard. The estimate shown by /stats model is Qwen Code’s calculation, not the amount charged. Check the actual amount in the connected provider’s usage records.

Limitations to know

• Built-in web_search is enabled or disabled depending on the connection target. Because it uses DashScope server-side search, it is enabled for ModelStudio Standard API Key and Token Plan entries and entries pointing to recognized DashScope hosts; it is disabled for Coding Plan (unverified on that endpoint) and for custom endpoints from third parties or other hosts (web_search documentation). Attach an MCP search server if needed.
• The Kunavo route is chat-only. Kunavo has no embedding, text-to-speech, or speech-recognition models. Qwen Code’s Live Voice requires a DashScope endpoint, so it uses a separate key unchanged regardless of where you route the chat model.

Payment when testing with Kunavo

Kunavo uses prepaid top-ups with no monthly fee; charges are deducted from your balance by token. The minimum top-up is $10, and Stripe Checkout accepts cards (Visa, Mastercard, American Express, JCB, UnionPay), Apple Pay, Google Pay, and Link. If Checkout is displayed in Korean won, you may also choose KakaoPay, Naver Pay, PAYCO, Samsung Pay, and domestic cards that cannot make international payments (Stripe's currency-conversion fee of 2–4% applies to conversion to won; paying in dollars avoids it). Toss Pay is not available. Review the payment information and create an account to issue a key. After your first request, the most reliable way to verify the actual amount is in your usage records.

Frequently asked questions

Checked October 1, 2026: the Qwen Code README; authentication, model provider, command, and web_search documentation (main branch); the repository’s UI translation file list; @qwen-code/qwen-code 0.24.7 on npm; and Alibaba Cloud’s Coding Plan and Token Plan pages. Kunavo has not run Qwen Code against its own endpoint.