Qwen Code was a terminal coding assistant. As of November 11, Alibaba repositioned it into something more interesting: an embeddable agent runtime. The production harness that powers Alibaba’s own Qoder — the machinery that plans multi-step tasks, orchestrates tool calls, and keeps an agent on track across a long job — is now open-sourced inside Qwen Code 1.0 Preview under Apache-2.0. This is not a CLI version bump. It is infrastructure.
What the Qoder Merger Actually Means #
Qoder was Alibaba’s proprietary AI-assisted development tool. It ran a production agent harness that the company used internally — handling task planning, tool orchestration, failure recovery, and state persistence across long sessions. That harness is now inside the open-source Qwen Code project.
The practical implication: Alibaba still ships Qoder using this exact same harness. What you are getting access to is production-proven agent infrastructure, not a demo. Every other terminal coding CLI — Claude Code, Codex, Aider — is a tool you use. Qwen Code 1.0 is a runtime you can build on top of.
import { createAgentSession } from '@qwenlm/qwen-code';
const session = await createAgentSession({
model: 'qwen-coder-plus', // or any OpenAI/Anthropic/local model
workdir: process.cwd(),
});
await session.run('Add TypeScript types to all exported functions');
That four-line embed is new. The stable 0.x branch did not expose this surface.
The Event Bus: Hand Separated from Brain #
The architectural decision in 1.0 is clean: reasoning (inference) and tool execution are decoupled through an event bus. In the preview, the bus runs in-process and in-memory. Events between the reasoning layer and the execution layer are first-class — observable, interceptable, replayable.
The public roadmap is direct about where this goes: in-process today, persistent event log next, remote agent execution after that, full multi-node distributed deployment at the end. If you build on the harness now, the event contract is designed to hold as the backend scales. You are not writing throwaway integration code.
What You Get: Auto-Memory, Auto-Skills, Agent Teams #
Beyond the new harness, Qwen Code ships a full agent development stack:
- Auto-Memory : SQLite-backed persistence for conversation history, tool results, and user preferences. Sessions resume where they left off.
- Auto-Skills : If the agent detects the same multi-step workflow three times, it proposes a new named skill. Skills accumulate over time.
- Agent Teams : A leader agent spawns named parallel teammates, delegates subtasks, and consolidates a final report. Useful for large refactors.
- MCP : Connect external tools, databases, and APIs through the Model Context Protocol. Zero extra setup required.
- Non-interactive mode : Session resumption and scripted execution for CI pipelines.
No Vendor Lock-In, By Design #
Qwen Code supports OpenAI, Anthropic, Gemini, and Qwen APIs out of the box, plus Ollama and vLLM for local models. You can switch models at runtime without restarting a session. This is a direct contrast to the two most popular alternatives:
| Tool | License | Model Support | Embeddable Harness |
|---|---|---|---|
| Claude Code | Proprietary | Anthropic only | No |
| Codex CLI | Open Source | OpenAI only | Partial |
| Qwen Code 1.0 | Apache-2.0 | Any | Yes |
For teams already running multiple models or self-hosting on Ollama, this removes the switching cost entirely.
Caveats Worth Flagging #
The 1.0 event bus is opt-in; the stable branch is retained separately. The distributed features — remote execution, multi-node — are roadmap items, not shipped code. The harness is production-proven inside Qoder’s closed-source form, but the open-source preview needs community testing before you bet a production system on it.
Use it. Evaluate it. The community breakdown on DEV.to and the full CLI comparison guide are good starting points for calibrating where it sits in your stack.
The repository is at github.com/QwenLM/qwen-code. Apache-2.0. No accounts required.