QwenLM/qwen-code

Qwen Code is an open-source AI coding agent, originally derived from Google's Gemini CLI v0.8.2, that works inside a project directory and answers requests such as explaining a repository or locating where to start. It is aimed at developers who want an agentic coding assistant in the terminal and are willing to supply a model provider and API key.

What changes in the workflow

The core loop is conversational but acts on files. You start qwen in a project, configure a provider with /auth, then issue a natural-language request Quick Start. The agent bundles Auto-Memory, Auto-Skills, SubAgents, Agent Teams and MCP, plus a set of built-in skills such as /review, /batch and /bugfix Capabilities. In practice, this means code review and batch edits can be requested as commands rather than assembled from scratch, and context from earlier sessions can be carried forward within the project.

Ways to use it

The terminal is the primary path, but the same agent is exposed through several interfaces interface table:

• Desktop app for macOS, Windows and Linux, downloaded from a release tag.
• Web UI via qwen serve --open, marked experimental.
• Editor plugins for VS Code, Zed and JetBrains.
• Chat channels for Telegram, DingTalk, WeChat and Feishu.
• Headless mode (qwen -p) for scripts and CI, plus TypeScript, Python and Java SDKs and an experimental daemon over HTTP + SSE.

Each path has its own setup page; the interface overview is the sensible first stop user overview.

Requirements and limits

• The npm route requires Node.js 22 or newer; there are also standalone shell/PowerShell installers and a Homebrew formula Installation.
• A model provider and API key are required. The agent supports OpenAI, Anthropic, Gemini and Qwen APIs, plus third-party providers and local models through Ollama or vLLM, switchable at runtime Authentication.
• The Web UI and daemon are explicitly labelled experimental, so they should not be treated as stable interfaces interface table.
• The project's own evaluation used the same Qwen 3.7 Max model across seven Qwen Code versions, each run on 500 SWE-bench Verified tasks with three trials per version, yielding averages of 76.40%–77.80% Evaluation. This measures the agent-plus-model pairing on that benchmark, not general coding reliability; full configuration is in the evaluation document.
• The framework and the Qwen models are open-source under Apache-2.0, but the README also notes that the agent is being used on its own repository for issues, pull requests, reviews and tests, and data handling for the Chrome extension and Browser Use is covered separately in the privacy notice Privacy.

HotRepo's take

The only recorded daily-board observation is 28,827 stars with +40 on 2026-10-30, which is far too little history to read as a trend repository. The HN discussion from July 2025 raised semantic drift over longer multi-turn sessions as a concern; that is an anecdotal observation about an early version rather than a current measured result HN thread.