qwen2api

Qwen-Proxy is a proxy service that converts https://chat.qwen.ai and Qwen Code / Qwen Cli into OpenAI-compatible APIs. It allows you to use any OpenAI-compatible client (such as ChatGPT-Next-Web or LobeChat) to call various Qwen models using a single account.

Key Features:

• OpenAI Compatibility: Seamlessly integrates with OpenAI-compatible clients.
• Anthropic Compatibility: Supports Anthropic Messages API (/v1/messages) for clients like Claude Code and the Anthropic SDK.
• Function Calling: Supports OpenAI tools and Anthropic tools, including streaming arguments incremental chunks and tool_choice=required retry logic.
• Multi-Account Support: Supports multi-account polling to improve availability.
• Advanced Capabilities: Supports streaming/non-streaming responses, multi-modal inputs (image/video recognition and generation), intelligent search, and deep thinking.
• CLI Endpoint: Provides a /cli endpoint with 256K context and tool-calling capabilities via Qwen Code / Qwen Cli.

The project is divided into a backend (src/) and a frontend (public/).

Backend (src/) Key Components:

• server.js: The main entry point for the server.
• start.js: An intelligent startup script that automatically detects whether to run in single-process or multi-process mode.
• config/: Contains configuration logic.
• controllers/: Handles core logic for chat, image/video generation, CLI chat, and models.
• routes/: Defines API endpoints for accounts, chat, CLI, models, settings, and verification.
• middlewares/: Includes authorization and chat-specific middleware.
• utils/: Contains critical utilities such as account-rotator.js (for account rotation), cookie-generator.js, fingerprint.js (for browser fingerprinting), and proxy-helper.js.
• data/: Stores persistent data in data.json.

Frontend (public/) Key Components:

• src/: Vue.js source code including views for auth.vue, dashboard.vue, and settings.vue.
• dist/: The directory containing compiled production assets.

Infrastructure & Deployment:

• docker/: Contains Dockerfile and docker-compose.yml for containerized deployment.
• ecosystem.config.js: PM2 configuration file for process management.

To prevent mass bans caused by IP correlation, you can assign a specific outgoing proxy to each individual account. The priority for proxy selection is: account.proxy > global PROXY_URL > no proxy.

Supported protocols: HTTP, HTTPS, SOCKS5.

Methods of Configuration:

1. Via Dashboard (Recommended): Open the dashboard, fill in the "Proxy URL" field when adding an account, or click "Change proxy" on an existing account card.
2. Via Environment Variables (when DATA_SAVE_MODE=none):
   • Old Format: ACCOUNTS=user1@mail.com:pass1,user2@mail.com:pass2 (no proxy specified).
   • New Format: Use | to separate the proxy URL. You can mix old and new formats.

Example (New Format):

ACCOUNTS=user1@mail.com:pass1|http://10.0.0.1:8080,user2@mail.com:pass2|socks5://10.0.0.2:1080

Example (JSON data/data.json in file mode):

{
  "accounts": [
    {
      "email": "user@mail.com",
      "password": "...",
      "token": "...",
      "expires": 1234567890,
      "proxy": "http://10.0.0.1:8080"
    }
  ]
}

⚠️ Warning: Proxy URLs in API responses are not masked. This project is intended for use in trusted local/private networks with a single administrator.

# New format (URL proxy via `|`, can be mixed with old format)
ACCOUNTS=user1@mail.com:pass1|http://10.0.0.1:8080,user2@mail.com:pass2|socks5://10.0.0.2:1080

The /v1/chat/completions endpoint supports the full OpenAI Function Calling protocol. Even if the upstream web interface lacks native tool capabilities, this service simulates them via prompt injection and streaming state machine parsing.

Key Features:

• Token Efficiency: Automatically compresses tools[] into TS-style signatures in the prompt, saving ~70% tokens.
• Streaming Support: Follows OpenAI spec for chunking function.name and arguments.
• Tool Choice: Supports "auto", "required", "none", and specific function selection.
• Automatic Retry: If tool_choice="required" is set but no tool is triggered on the first attempt, the service automatically appends a constraint hint to retry.

{
  "model": "qwen3-coder-plus",
  "stream": true,
  "messages": [
    {"role": "user", "content": "Check the weather in Beijing"}
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get city weather",
        "parameters": {
          "type": "object",
          "properties": { "city": { "type": "string" } },
          "required": ["city"]
        }
      }
    }
  ],
  "tool_choice": "required"
}

Qwen2API provides an Anthropic-compatible bridge at the /v1/messages endpoint. This allows tools like Claude Code, Anthropic SDK, and aider to work with Qwen models.

Important Compatibility Notes: Qwen2API is a compatibility layer, not a native Anthropic backend. Some fields are implemented via 'soft' strategies (prompt injection or simulation) rather than native upstream semantics.

• Supported: model, messages.text, messages.image, stream.
• Partial Support: messages.tool_use, messages.tool_result, system (injected as prompt prefix), messages (multi-turn history compression), tools[] (via prompt/XML simulation), tool_choice (via prompt steering), thinking (legacy format support).
• Ignored with Warning: max_tokens, stop_sequences, metadata, temperature, top_p, top_k, service_tier, container, output_config. These fields will not affect upstream behavior but will trigger warnings in response headers.
• Not Supported: mcp_servers, context_management.

If unsupported or partially supported fields are used, the API may return these headers:

• X-Qwen2API-Anthropic-Compatibility
• X-Qwen2API-Anthropic-Warnings