⌘K
Guides

Use AI Hub with Qwen Code

Qwen Code is an open-source coding agent for the terminal. It runs on AI Hub through its OpenAI-compatible protocol: set the base URL and your API key, and pick any catalogue model.

1. Install

Qwen Code needs Node.js 22 or later.

bash
npm install -g @qwen-code/qwen-code@latest

2. Point it at AI Hub

The quickest setup is three environment variables. OPENAI_MODEL takes a catalogue ID in author/slug form.

bash
export OPENAI_API_KEY="sk-tokenhub-…"
export OPENAI_BASE_URL="https://tokenhub-api.belive.sg/api/v1"
export OPENAI_MODEL="qwen/qwen3-coder-flash"
qwen

To switch between several AI Hub models with /model, declare them in ~/.qwen/settings.json. The key stays in the environment variable named by envKey; Qwen Code never writes it to the file.

json
{
  "modelProviders": {
    "openai": [
      { "id": "qwen/qwen3-coder-flash", "name": "Qwen3 Coder Flash (AI Hub)",
        "baseUrl": "https://tokenhub-api.belive.sg/api/v1", "envKey": "TOKENHUB_API_KEY" },
      { "id": "deepseek/deepseek-v4-pro", "name": "DeepSeek V4 Pro (AI Hub)",
        "baseUrl": "https://tokenhub-api.belive.sg/api/v1", "envKey": "TOKENHUB_API_KEY" },
      { "id": "moonshotai/kimi-k3", "name": "Kimi K3 (AI Hub)",
        "baseUrl": "https://tokenhub-api.belive.sg/api/v1", "envKey": "TOKENHUB_API_KEY" }
    ]
  },
  "security": { "auth": { "selectedType": "openai" } },
  "model": { "name": "qwen/qwen3-coder-flash" }
}

Use the openai protocol for the Qwen Code recipe here. See Messages and Responses for AI Hub's other protocols.

3. Run it

Run qwen in a project directory for the interactive agent, or pass a prompt for a one-shot run in a script or CI job:

bash
qwen -o json "Fix the failing test in test_prices.py and run the tests"

Every turn is an ordinary streamed chat completion, so each one appears in your Activity log with its model, tokens and cost.

Choose a model

Choose an available model from Models whose supported_parameters includes tools. Use its exact ID in the client settings and check its context limit before opening a large project.

Cost

  • Every turn resends Qwen Code's system prompt, tool definitions and conversation history. Input usage grows with the project context and the number of turns.
  • Most of that input repeats the previous turn. Models whose upstream reports cached input show it in usage.prompt_tokens_details.cached_tokens.
  • Give the agent its own key with a monthly spend cap, so a runaway session stops at the cap. See Rate limits & caps.
← PreviousCostNext →Hermes Agent