Qwen Code: Claude & GPT via OpenAI-Compatible

Qwen Code is Alibaba's open-source terminal agent; its modelProviders setting accepts any OpenAI-compatible base URL, so one Code By AI key serves Claude and GPT.

Qwen Code is a terminal coding agent derived from Gemini CLI. It reads ~/.qwen/settings.json and can register several OpenAI-compatible models, which the /model picker switches between.

One command (macOS / Linux)

Copy
1
curl -fsSL https://www.codebyai.net/setup.sh | sh -s -- --qwen-code

It asks for your API key (hidden input), writes CODEBYAI_API_KEY to your shell rc file, and adds gpt-5.6-terra and claude-sonnet-5 to ~/.qwen/settings.json as the default models, backing the file up first. The key never goes into settings.json. With QWEN_HOME set, it edits the file there, as Qwen Code does. Re-running rotates the key and keeps the model you picked with /model; | sh -s -- --uninstall restores the file. Open a new terminal, then run qwen. Comments in settings.json do not survive the rewrite; the backup keeps them. The manual steps below do the same by hand.

1. Install Qwen Code

Node.js 22 or newer:

Copy
1
npm install -g @qwen-code/qwen-code@latest

On macOS or Linux, brew install qwen-code is an alternative. Check the install with qwen --version.

2. Create a Code By AI API key

Sign in to Code By AI, open API Keys, and create a key. It looks like sk-proj-....

3. Add the models to settings.json

Open ~/.qwen/settings.json (user level) and merge this into the existing object. Keep any providers already there:

Copy
1
{
2
  "modelProviders": {
3
    "openai": [
4
      {
5
        "id": "gpt-5.6-terra",
6
        "name": "CodeByAI GPT",
7
        "baseUrl": "https://www.codebyai.net/api/agent/v1",
8
        "envKey": "CODEBYAI_API_KEY"
9
      },
10
      {
11
        "id": "claude-sonnet-5",
12
        "name": "CodeByAI Claude",
13
        "baseUrl": "https://www.codebyai.net/api/agent/v1",
14
        "envKey": "CODEBYAI_API_KEY"
15
      }
16
    ]
17
  },
18
  "security": { "auth": { "selectedType": "openai" } },
19
  "model": { "name": "gpt-5.6-terra" }
20
}

The openai provider ID means the OpenAI-compatible protocol. wireApi is omitted, so Chat Completions is used; the gateway serves both model families on that route. Do not add wireApi: "responses" unless you deliberately want the Responses API.

4. Export the key

Credentials are read from the environment, never stored in the settings file:

Copy
1
export CODEBYAI_API_KEY="sk-proj-xxxx"

Add that line to ~/.zshrc or ~/.bashrc, then start a new shell. If the variable is missing, Qwen Code skips the route or fails authentication instead of falling back silently.

5. Start Qwen Code

Copy
1
cd your-project
2
qwen

Run /model, pick gpt-5.6-terra or claude-sonnet-5, and send a prompt. The picker stores the choice for the session.

Troubleshooting

• The models do not appear in /model — modelProviders is not a bare array, or the provider ID is not a built-in auth type. Use "openai": [ ... ] as shown.
• 401 — CODEBYAI_API_KEY is not exported in the shell that started qwen. echo $CODEBYAI_API_KEY should print the key.
• 404 — baseUrl includes /chat/completions or duplicates /v1. Use https://www.codebyai.net/api/agent/v1.
• Only one family works — confirm the model ID matches the Model Plaza; claude-* is forwarded to the Claude channel, everything else is treated as a Codex-channel model.