How to run an AI agent that fixes code and runs tests in your terminal
Once you enter a project folder and type the qwen command, you can watch the AI analyze the repository structure, modify necessary files, and execute tests. This is an open-source AI coding agent that operates across various environments, including terminals, desktops, browsers, and chat apps. The Qwen 1.0 Preview release combines the agent capabilities of Qoder and Qwen Code into a single offering.
What changed in the 1.0 Preview
This 1.0 Preview is an open-source release that integrates the agent harness features of Qoder and Qwen Code. Automatic memory, skills, sub-agents, and MCP work by default from the first session. Agent teams, agent arenas, and dynamic workflows are included as experimental or optionally enabled features. Both the framework and Qwen models are open source, allowing them to evolve together without vendor lock-in.
Workflows that continue across different environments
You can start in the terminal and switch to a desktop app or web UI. Editor plugins for VS Code, Zed, and JetBrains are also supported. You can connect to chat channels like Telegram, WeChat, and GitHub to issue remote commands. Automation scripts or CI/CD pipelines can run in headless mode, and custom integrations can be built using TypeScript, Python, and Java SDKs.
How the agent remembers and extends itself
The agent has an automatic memory feature that writes and reloads its own notes between sessions. Instructions in QWEN.md or AGENTS.md files are loaded automatically. You can use built-in skills like /review and /bugfix directly, or create new skills by analyzing documentation or code with the /learn command. You can connect MCP servers to add tools and resources, and set up hooks to run your own scripts before or after tool calls.
Model selection and performance verification
It supports OpenAI, Anthropic, Gemini, and Qwen APIs, and local models like Ollama or vLLM can be switched at runtime. On the SWE-bench Verified benchmark, the Qwen 3.7 Max model recorded an average score between 76.40% and 77.80% across 7 versions. This result comes from running 500 tasks three times per version.
Things to check before use
When running in daemon mode, the default configuration listens only on the loopback address and does not require authentication. On shared hosts or in untrusted environments, you must set QWEN_SERVER_TOKEN and use the --require-auth option. Node.js 22 or higher is required for NPM installation. You must restart your terminal after installation for environment variables to take effect. This repository started based on Google Gemini CLI v0.8.2 but has been developed independently since v0.1.