Qwen Code Draft
Qwen · reviewed 22 November 2026 · its card
Qwen Code helps a developer "turn ideas into code faster than ever before": "Qwen's agentic coding tool", "the open-source AI coding agent for your terminal, editor, desktop, browser, and chat" (S1, S2). It is "maintained by the Qwen Code team" in the QwenLM GitHub organisation, and was "originally based on Google Gemini CLI v0.8.2" (S1, S22); its own pages do not name Alibaba, which secondary coverage gives as the maker (S25), and Alibaba Cloud's Model Studio documents it for its models (S23). It runs as a terminal CLI, headless, a desktop app for macOS, Windows and Linux, a browser Web Shell from qwen serve, in VS Code, Zed and JetBrains, through chat channels, and through TypeScript, Python and Java SDKs (S1, S20). It is Apache-2.0 open source, with copyright lines for Google LLC and Qwen (S5). The latest stable release seen is v0.28.3, of 21 November 2026; a Qwen Code 1.0 Preview pre-release "brings together the agent capabilities of Qoder and Qwen Code" (S3, S4). Qoder, "an agentic coding platform" with desktop, mobile, IDE, JetBrains and CLI products, is the other agentic coding product in that line; its page does not name its maker, and secondary coverage calls it "Alibaba's agent-assisted development tool" (S24, S25). It is an agent harness and a client.
- An instance of
- Agent harness, Client
Capabilities
| Capability | Value | Detail | Sources |
|---|---|---|---|
| Model | Yes | "Supports OpenAI, Anthropic, Gemini, and Qwen APIs. Any third-party provider or local model (Ollama / vLLM)"; Qwen account sign-in, default qwen3.5-plus; Alibaba's Coding Plan adds Qwen, GLM, Kimi and MiniMax models; Qwen4-Max integrated |
S1 S18 S21 |
| Session | Yes | "Resume past conversations with /resume, and browse, rename, or archive sessions"; /branch forks a conversation into a new session; qwen sessions list |
S1 S16 |
| Subagent | Yes | Markdown agents in .qwen/agents/ and ~/.qwen/agents/; "Top-level regular subagents run in the background by default"; "Fork children cannot spawn any further sub-agent" |
S6 |
| Agent team | Partial | "The experimental Agent Team runtime": teammates "share a task list, exchange messages"; a leader assigns up to three workstreams; teammates are in-process, not separate sessions | S15 |
| Skill | Yes | SKILL.md folders from .qwen/skills/, .agents/skills/, global folders, extensions and bundled skills; the page does not name the specification |
S7 |
| Instruction file | Yes | QWEN.md in ~/.qwen/ and the project root, .qwen/QWEN.local.md; "If your repository already has an AGENTS.md file for other AI tools, Qwen reads that too" |
S8 |
| Plugin | Yes | Extensions "package prompts, MCP servers, subagents, skills and custom commands", from Git, npm, archives or a marketplace; "Claude plugins are automatically converted" | S9 |
| MCP | Yes | Servers over stdio, Streamable HTTP and legacy SSE; "supports OAuth 2.0 authentication for MCP servers" | S10 |
| Execution environment | Yes | harness-sandbox or container, opt-in: macOS Seatbelt, Linux bwrap or Landlock for tools, Docker or Podman; "Sandboxing reduces but doesn't eliminate all risks" |
S11 |
| Client | Yes | Terminal, desktop app, browser Web Shell, VS Code, Zed and JetBrains, chat channels; phones through the Web Shell over TLS | S1 S19 S20 |
| Agent host | Partial | Built-in claude-code and codex agents "delegate to separately installed native tools", Claude Code through its claude-agent-acp adapter; "Native sessions cannot be restored after restarting Qwen Code" |
S6 |
| Channel | Yes | Telegram, WeChat, DingTalk, WeCom, Feishu, QQ, Email, GitHub, GitLab and webhooks; it "routes incoming messages to the agent and sends responses back to the correct chat"; the daemon-managed mode is experimental | S13 S19 S3 |
| Cross-session messaging | Yes | Running sessions are "addressable from send_message" over one UNIX socket each, by name or ref; an ACP-driven session refuses |
S14 |
| ACP | Yes | As agent: qwen --acp in Zed, installed "from ACP Registry"; as client: it runs Claude Code through "the ACP executor" |
S12 S6 |
| AHP | Unknown | Not found | None |
| A2A | Unknown | Not found; the docs list no A2A page, and cross-workspace delegation is not named A2A | S2 S21 |
| Bot | Unknown | Not found. Auto-memory in ~/.qwen/projects/<project>/memory/; durable scheduled tasks under channels and the daemon; the experimental Managed Agent deployment runs sessions by API call |
S8 S17 S26 |
State of play
How things stand, newest first, as the card records it.
- Version 0.28.3, the latest release: GitHub becomes a managed channel adapter beside a generic webhook channel, and ACP sessions run on the shared agent loop by default. S3
- The weekly update covers v0.27.3 to v0.28.0: Qwen4-Max is integrated and new Web Shell sessions get their own worktree by default. S21
- Version 0.28.0 and Qwen Code Desktop v0.28.0. S3
- Qwen Code 1.0 Preview, a pre-release that "brings together the agent capabilities of Qoder and Qwen Code" and "supports Qwen and Qoder accounts"; the tag page gives no year, and the year and its being "a separate branch" come from secondary coverage. S4 S25
- The desktop app and Web Shell blog post: "Both do the same work and look the same." S20