Set up Qwen Code: Installation, Providers, German Interface, and API

Qwen Code is the open-source AI coding agent (Apache-2.0) from Alibaba’s Qwen team. It is set up in four steps: install it, start it in the project with qwen, choose the provider with /auth, and give it a task. Important beforehand: the free Qwen OAuth allowance ended on April 15, 2026 — anyone following an older guide can no longer continue for free that way. This page follows the current /auth menu, shows how to switch to German, explains how to configure a custom endpoint for Claude or GPT, and covers the most important commands. The latest version is v0.24.7, released on npm on September 29, 2026; releases arrive at least weekly, so check your version with qwen --version.

Installation

The commands come from the official README. With the standalone installer, you do not need to provide Node.js yourself; only npm requires Node.js 22 or newer.

# macOS / Linux (offizieller Standalone-Installer)
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 (benötigt Node.js 22 oder neuer)
npm install -g @qwen-code/qwen-code@latest

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

Reopen the terminal afterward so the environment variables take effect. Qwen Code was originally based on Google Gemini CLI v0.8.2, but has been developed independently since v0.1 and is no longer synchronized with the original. Gemini CLI settings and quotas do not apply here.

First launch and German interface

cd /pfad/zu/ihrem-projekt
qwen

# in der Sitzung:
/language ui de-DE        # Oberfläche auf Deutsch
/language output German   # Antworten des Modells auf Deutsch
/auth                     # Anbieter und API-Key einrichten

After /language ui de-DE, Qwen Code displays its own German labels, such as “Authentifizierungsmethode auswählen” or “Werkzeug-Genehmigungsmodus”; this page uses the same terms. The response language is separate and is set with /language output German.

Besides the terminal, there is a desktop app, a web interface in the browser (qwen serve --open, experimental), integrations for VS Code, Zed, and JetBrains, and headless mode for scripts and CI (qwen -p "..."). Everything is in the same free repository; inference is billed separately.

Choose the provider with /auth

/auth (alias /login) opens “Authentifizierungsmethode auswählen” with three top-level options. Anyone who still tries to choose Qwen OAuth sees: “Discontinued — switch to Coding Plan or API Key”.

The three ModelStudio options are not three payment methods for the same bill: the documentation gives each its own endpoint and key, and the key type and baseUrl must match. On October 1, 2026, the Coding Plan was listed as “limited availability, first come first served, replenished daily at 00:00 (UTC+8)” — so it may not be purchasable on a given day. A task consumes multiple model calls on the Coding Plan; Alibaba typically cites 5–10 for simple tasks and 10–30 or more for complex tasks. The English page Qwen Code pricing covers the cost comparison in detail.

Use Claude or GPT: Custom Provider in settings.json

The authentication documentation recommends the modelProviders block in ~/.qwen/settings.json for third-party providers such as OpenAI, Anthropic, Google, OpenRouter, or your own endpoint. For Kunavo, it looks like this:

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

# Den Key nicht in settings.json schreiben, sondern in .qwen/.env oder die Umgebung
echo 'KUNAVO_API_KEY=sk-kn-...' >> ~/.qwen/.env

Four points that save troubleshooting time:

• baseUrl ends at /v1. The Model Providers reference explicitly says to set it to the API’s /v1 root, not /v1/chat/completions — the SDK appends the path itself. Adding the path results in a 404, not an authentication error.
• Where the key is stored. Qwen Code reads it from the environment variable named by envKey. Order: shell-export, then a .env file (.qwen/.env recommended; only the first file found counts), and finally env in settings.json — the key is stored there in plain text.
• The file overrides the CLI option. Order: values from /auth → selected modelProviders entry → CLI arguments → environment variables → settings.json. That is why --openai-base-url sometimes appears ineffective. security.auth.apiKey and security.auth.baseUrl from older guides are outdated.
• Start with Chat Completions. Without wireApi, Chat Completions is used. With "responses", there is neither endpoint detection nor automatic fallback. Changes to modelProviders also take effect in an active session (/model reopen it).

There is no Qwen text model in the Kunavo catalog. So this is not a cheaper route to Qwen, but a way to use Claude or GPT in Qwen Code with prepaid credit. Kunavo also did not run Qwen Code itself against its endpoint; like the English Qwen Code setup page, this configuration comes from the official documentation. Keep a working route while you try it.

First task and important commands

The README essentially suggests this as the first task: “Explain this repository and show me where to start.” It checks file access and tool calls at once — a simple “Hello” does not expose connection problems.

For yolo and the other automatic modes, the documentation itself warns: use them only in trusted, isolated, or disposable environments. The estimated costs in /stats model are calculated by Qwen Code itself; the actual amount appears in your provider’s usage log.

Limits you should know

• Built-in web search depends on the host. It uses DashScope’s server-side search: enabled for the ModelStudio Standard API Key and Token Plan, as well as entries on a detected DashScope host; disabled for the Coding Plan (not verified there) and for third-party providers and custom endpoints on other hosts (web_search documentation). Alternative: an MCP search server.
• Kunavo only provides chat. Kunavo does not offer embedding, text-to-speech, or speech-recognition models. Qwen Code’s Live Voice requires a DashScope endpoint anyway and continues using its own key, regardless of where the chat model points.

Paying when you try Kunavo

Kunavo uses prepaid balance with no subscription and charges per token. The minimum top-up is $10; the Stripe checkout offers cards (Visa, Mastercard, American Express), Apple Pay, Google Pay, and Link, plus EPS in Austria — SEPA Direct Debit is currently not offered. Stripe offers Pay by Bank only to buyers in the United Kingdom, Ireland, and Finland; in Germany it is currently available only as a closed preview (Stripe documentation, as of October 3, 2026). See Billing for details; create the key with an account. After the first task, the usage log shows the actual amount.

Frequently asked questions

Checked on October 1, 2026: the Qwen Code README; the authentication, Model Providers, commands, and web_search documentation (main branch); the German UI text (packages/cli/src/i18n/locales/de.js); @qwen-code/qwen-code 0.24.7 on npm; and Alibaba Cloud’s Coding Plan and Token Plan pages. Kunavo did not run Qwen Code against its endpoint.