Qwen Code is an open-source agentic CLI coding assistant released by Alibaba's QwenLM team on July 22, 2025, forked from Google's Gemini CLI and optimised for the Qwen3-Coder family of models. Launched alongside the Qwen3-Coder-480B model, it accumulated over 25,000 GitHub stars and 2,500 forks rapidly, reflecting strong developer demand for an Apache 2.0-licensed alternative in the agentic coding space. It supports Claude, GPT-4o, Gemini, and other major providers alongside its native Qwen backends.
Background
Qwen Code originates from Alibaba's Tongyi Lab, the AI research division of Alibaba Group that has operated under various names since the founding of the DAMO Academy in October 2017. The Qwen series of language models (short for Tongyi Qianwen, meaning "seeking the truth of the thousand questions") launched publicly in September 2023 after an April 2023 beta. Over the following two years, the team released Qwen2 in June 2024 and Qwen3 in April 2025, each generation extending multilingual coverage and benchmark performance. By May 2026, the Qwen app had reached 234 million users globally, with Qwen-based coding models surpassing 20 million downloads.
Qwen Code itself was announced on July 22-23, 2025, packaged as the companion CLI for the newly released Qwen3-Coder models. Alibaba explicitly acknowledged the fork origin in the repository README: "This project is based on Google Gemini CLI," inheriting its TypeScript-based architecture (approximately 92% TypeScript) while adding parser-level adaptations for Qwen-Coder models and substantially broadening provider support. The project carries the Apache 2.0 license, matching Gemini CLI, and places no restrictions on commercial use or modification. It has shipped over 480 releases since launch, with v0.18.0 released in June 2026.
Key capabilities
Qwen3-Coder model optimisation and free access tier
The tool's native backend is Qwen3-Coder, specifically the flagship Qwen3-Coder-480B-A35B-Instruct variant: a 480-billion-parameter Mixture-of-Experts model with 35 billion active parameters per token. It was trained on 7.5 trillion tokens at a 70% code ratio, supporting a 256K-token context window natively and up to 1 million tokens via extrapolation. On SWE-Bench Verified — the standard benchmark for autonomous software engineering — it achieved state-of-the-art results among open-source models, comparable to Claude Sonnet 4 according to Alibaba's announcement. Access to the model is available for free through Alibaba Cloud's Coding Plan and via the Qwen Chat platform, reducing the barrier for individual developers and small teams who want to evaluate the tool without incurring API costs.
Gemini CLI foundation and agentic architecture
Qwen Code retains Gemini CLI's core planner-executor design, in which the model receives a task, issues tool calls (file reads, edits, terminal commands, git operations), observes results, and iterates toward completion. The tool loop runs at autonomy level 3: it can handle multi-step coding tasks — reading context, proposing edits, running tests, and correcting on failure — but by default presents diffs and awaits confirmation before writing changes to disk. Operators can adjust approval thresholds. The tool also supports a daemon mode (qwen serve) that exposes an HTTP and Server-Sent Events interface, enabling shared sessions and programmatic integration from external scripts or CI pipelines. SDKs for TypeScript, Python, and Java wrap the daemon API.
VS Code and Zed integration
Beyond the standalone terminal, Qwen Code ships a VS Code Companion extension (available in the VS Code Marketplace as "Qwen Code Companion") that embeds a React-based chat panel directly in the editor. The extension acts as an MCP server exposing IDE capabilities — file system access, open editors, terminal output — back to the Qwen Code CLI process. Every server instance generates a unique UUID auth token, and all requests must supply it as a Bearer header, preventing unauthorised local access. JetBrains IDEs are supported through the ACP registry, and Zed integration follows the same model with native Zed extension hooks. GitHub Actions integration is also documented for CI workflows.
MCP compatibility
Qwen Code implements the Model Context Protocol as both a client and a host. As a client it can connect to any compliant MCP server — including community servers for Google Drive, Figma, Slack, databases, and internal APIs — using three transport mechanisms: stdio (subprocess), SSE (Server-Sent Events), and streamable HTTP. The configuration in settings.json supports per-server tool filtering (includeTools/excludeTools), trust levels (bypass confirmation dialogs for known-safe servers), and OAuth 2.0 with automatic discovery, dynamic client registration, and token persistence in ~/.qwen/mcp-oauth-tokens.json. MCP servers can also expose predefined prompts as slash commands within the Qwen Code session, enabling repeatable workflow shortcuts without modifying the core tool.
Autonomy level
Qwen Code operates at autonomy level 3. It can plan and execute multi-step coding tasks — reading context, proposing edits, running tests, and iterating — but presents diffs and awaits confirmation before writing changes by default. Operators can adjust approval thresholds, and the tool will loop on failures, but it does not self-deploy or take irreversible production actions without explicit human sign-off.
Strengths
- Apache 2.0 license with no usage restrictions on the core tool or its extensions
- Native optimisation for Qwen3-Coder, with a free-tier API path via Alibaba Cloud Coding Plan and Qwen Chat
- Broad provider support (Claude, GPT-4o, Gemini, OpenRouter, Fireworks AI) for teams that are not committed to one model
- VS Code Companion extension with UUID-gated local MCP server for secure in-editor use
- Zed and JetBrains integrations reduce context-switching without locking users into a single IDE
- Full MCP client support with OAuth 2.0, three transport types, and slash-command shortcuts
- Daemon mode with HTTP/SSE API and TypeScript, Python, and Java SDKs for programmatic integration
- Active release cadence (483 releases since July 2025, v0.18.0 in June 2026) with a fast-growing community
Limitations
- Relatively new tool with a shorter production track record than Gemini CLI, Aider, or Claude Code
- Some features and prompt tuning are optimised for Qwen backends; behaviour with other providers may differ from documented examples
- English documentation is thinner than established competitors; a significant portion of issues, discussions, and community resources are in Chinese
- Forked lineage means some upstream Gemini CLI behaviours and assumptions may surface unexpectedly in edge cases until the codebase diverges further
- The 480B flagship model requires substantial infrastructure to self-host; teams without Alibaba Cloud access must rely on third-party API providers for the highest-capability tier