Qwen Code

What it is

QwenLM's open-source coding agent, and openly a fork: it takes the Gemini CLI codebase and adapts it for Qwen models. Apache-2.0, around 27,000 GitHub stars. The fork is not a secret or an embarrassment — the project states it, and the licence permits it — but it does shape what you get. The interface, the command set and the general feel will be immediately familiar to anyone who has used the Gemini CLI.

What changed is the fit. The prompts, the tool definitions and the expectations about how a model formats a tool call have all been retuned for Qwen. That is the whole value proposition: if you are running Qwen, this is the harness that was built to match it.

How the loop works

A conventional terminal agent loop in TypeScript. Read the repository, propose and apply edits, run commands, read the results, continue. Nothing here is architecturally novel, which is a compliment in context — the Gemini CLI loop is well-worn and the adaptation did not try to reinvent it.

Provider support goes wider than Qwen on paper: OpenAI, Anthropic and Gemini APIs are configurable, as are local runs through Ollama or vLLM. Local Qwen through vLLM is a particularly coherent setup, since both halves come from the same place.

Qwen offers a free tier through its own model access, which makes this one of the cheaper ways to try a full coding agent. Check the current terms before you plan around it — free tiers in this category have a habit of ending, as the Gemini CLI's did in June 2026.

Isolation and permissions

Configurable, in the way inherited from its parent codebase: you decide how much the agent may do without asking. There is no built-in sandbox, so containment is your job if you want it. Sensible practice is the same as for any terminal agent — run it in a working tree you are willing to lose, and keep secrets out of the environment it inherits.

Who it is for

People running Qwen models. That is the honest answer. If Qwen is your model — by preference, by cost, by a data-residency constraint, or because you are self-hosting it — Qwen Code is the best-fitting harness available and the pairing is a good one.

Limits

Take the multi-provider support with caution. The tool-call format expectations are tuned to how Qwen emits calls, and on other providers that tuning works against you: parsing gets brittle, malformed calls slip through, and results degrade in ways that look like model failures but are not. It is the known quirk of this harness as of Sep 2026. If your model is not Qwen, use a harness that was not built around one.

Being a downstream fork also means upstream Gemini CLI changes arrive on QwenLM's schedule, not Google's.

Alternatives on this board

• Gemini CLI — the upstream codebase, now a paid decision.
• OpenCode — genuinely provider-agnostic, if you are switching models often.
• OpenClaw — MIT and model-agnostic, without the single-vendor tuning.

Sources

• QwenLM/qwen-code on GitHub
• QwenLM blog