Qwen Code
QwenLM · AI coding
GitHub stars
28,811Awaiting first sync
No measured star history yet
What is Qwen Code?
Qwen Code gives developers a coding assistant in the terminal that can read a project, make changes and run commands. It is not tied to one model vendor: you can use Qwen models, other major APIs or a local model. Approval modes let a team decide how much the agent may do on its own, from read-only planning to fully automatic runs.
Qwen Code is an open-source AI coding agent from the Qwen team that runs in the terminal. It edits files, runs commands, creates commits and answers questions about a codebase, with subagents, skills, hooks and MCP support. It works with Qwen, OpenAI, Anthropic and Gemini APIs, other providers and local models through Ollama or vLLM.
Core capabilities
- Connects to Qwen, OpenAI, Anthropic and Gemini APIs, or local models via Ollama and vLLM.
- Five approval modes, from read-only Plan to fully automatic YOLO.
- Use it in the terminal, a desktop app, VS Code, Zed, JetBrains or chat apps.
Agent features
Qwen Code can build features from plain-language descriptions, debug issues, answer questions about a codebase and create commits. It supports subagents, agent teams, skills, hooks, plan mode, LSP integration, git worktrees and MCP servers. QWEN.md files and auto-memory carry instructions and learned notes between sessions. The project began as a fork of Gemini CLI 0.8.2 and has been developed independently since version 0.1.
- AI coding
- Subagents
- MCP
- QWEN.md
Models, approvals and interfaces
The /auth command connects QwenCloud or Alibaba ModelStudio plans and API keys, third-party providers such as DeepSeek and OpenRouter, or a custom endpoint, including local servers. The earlier free Qwen OAuth tier ended on 15 April 2026. There are five approval modes: Plan, Ask Permissions, Auto-Edit, YOLO and the default Auto mode, where a classifier approves safe actions and blocks risky ones. Besides the terminal, there is a desktop app, an experimental web interface started with qwen serve, editor plugins and connectors for Telegram, DingTalk, WeChat and Feishu.
- Multi-provider
- Approval modes
- Desktop app
- IDE plugins
Install and licence
Qwen Code is licensed under Apache 2.0. It installs with a standalone script for Linux, macOS or Windows, with npm on Node.js 22 or later, or with Homebrew, and starts with the qwen command in a project folder. Headless mode runs qwen -p in scripts and CI, and SDKs for TypeScript, Python and Java let developers build on it. Model costs depend on the provider and plan you connect.
- Apache-2.0
- npm
- Homebrew
- Headless
Where it fits
Use cases
- 01
Plan a change before any code is touched
Run qwen in the project folder and type /plan with the task, such as refactoring a module. In Plan mode it only reads and analyses the code, then proposes the steps. When you agree, press Shift+Tab to switch to Ask Permissions or Auto-Edit and let it make the changes.
- 02
Teach the agent your project's rules
Write build commands and conventions in a QWEN.md file at the project root and commit it so the whole team shares it. Personal preferences go in ~/.qwen/QWEN.md or .qwen/QWEN.local.md. Qwen Code also reads an existing AGENTS.md, and its auto-memory keeps notes it learns across sessions.
- 03
Split large tasks across specialised agents
Define subagents with their own system prompt and allowed tools for jobs such as code review or testing. Each subagent keeps its own conversation history and works on its task until it finishes, while you follow its progress. For unattended jobs, run qwen -p in scripts or CI.