Qwen Code

Built in tools read and write files, run shell commands, keep a todo list, fetch pages and search the web, with search backed by DashScope on Qwen and Model Studio plans. MCP servers connect over stdio or HTTP at user or project scope, with OAuth sign in, per server tool include and exclude lists, and MCP prompts and resources. Pull request fixes push through the GitHub command line tool, review comments post inline on pull requests, and GitHub and GitLab channels pick up repository activity.

Sourceqwenlm.github.io/qwen-code-docs/en/users/features/mcpread 2026-11-08

Explore, general purpose and fork subagents run in the background by default, each with its own model and tool allowlist, and can be isolated in their own git worktree. Plan mode, goals, loops and batch commands structure longer work, and background sessions can be reattached later. Agent Teams splits a task into up to three workstreams with one writer pinned to a worktree, and Arena races several models on the same task, though both are marked experimental.

Sourceqwenlm.github.io/qwen-code-docs/en/users/features/sub-agentsread 2026-11-08

A QWEN.md context file rides along on every request, rules folders add guidance scoped to paths, and @ references pull specific files into a prompt. Web fetch and search bring in outside pages, and an experimental language server mode adds symbol awareness. No embeddings, vector index or semantic search over the codebase or company documents is offered, so the agent finds code by reading and searching files as it works.

Sourceqwenlm.github.io/qwen-code-docs/en/users/features/rulesread 2026-11-08

Five approval modes run from read only planning, through asking before every action and auto approving file edits, to an auto mode where a classifier screens each call and falls back to manual after three blocks, and a mode that approves everything. Allow, ask and deny rules decide individual tools and commands, with deny winning. Checkpoints taken before each tool call let restore and rewind undo file changes, and optional sandboxes run tools in Docker, Podman, macOS Seatbelt or Linux bubblewrap.

Sourceqwenlm.github.io/qwen-code-docs/en/users/features/approval-moderead 2026-11-08

Operators can place system settings that override user and project choices, enforcing the sign in method, allowed approval modes and permitted MCP servers, and project hooks need approval before they run. Sign in uses a Qwen account, API keys or Vertex credentials, prompts and code are not used for training by the tool, and usage statistics go to Alibaba Cloud unless switched off. The local server has no authentication unless a token is required, and there is no single sign on, user roles or security attestation.

Sourceqwenlm.github.io/qwen-code-docs/en/users/support/tos-privacyread 2026-11-08

OpenTelemetry export sends traces, logs and metrics such as token usage and tool call counts to any OpenTelemetry endpoint or a file, with Alibaba Cloud's monitoring service supported, though prompts are logged by default once telemetry is on. A stats command shows usage by model, tool and skill over days and months, sessions export as web pages, Markdown or JSON, and headless output carries model and tool statistics for each run.

Sourceqwenlm.github.io/qwen-code-docs/en/developers/development/telemetryread 2026-11-08

Auto memory is on by default and keeps markdown notes about the developer, their feedback, the project and outside references in a folder per project, with an index file and pinned notes, and each git worktree gets its own folder. Commands view, add, remove and consolidate memories, the files can be edited directly, and a cleanup pass runs once a day after enough sessions. Shared team memory in the repository and a Mem0 integration are optional, and the whole feature can be turned off.

Sourceqwenlm.github.io/qwen-code-docs/en/users/features/memoryread 2026-11-08

Everything runs on the developer's own machine under Linux, macOS or Windows, and local models served by Ollama, vLLM or LM Studio keep code and prompts on that machine entirely. Hosted Qwen models come from Alibaba Cloud endpoints in Beijing for China and Singapore for international users, and no hosted agent service sits between the developer and the model. A self hosted reference deployment for a managed agent exists but is experimental.

Sourceqwenlm.github.io/qwen-code-docs/en/users/configuration/model-providersread 2026-11-08

Skills for review, batch work, loops and bug fixing come built in alongside Explore, general purpose and fork subagents, and Auto Skill writes new skills from repeated work, marking them stale after 30 days. Extensions install from git, npm or a URL, and the tool loads extensions from the Gemini CLI gallery, plugins from the Claude Code marketplace and Qoder plugins, browsable from a Discover tab. There is no marketplace run by the Qwen team itself.

Sourceqwenlm.github.io/qwen-code-docs/en/users/extension/introductionread 2026-11-08

Beyond the interactive terminal, headless runs take a prompt with turn and wall time limits and return text, JSON or streamed JSON for scripts and CI. A GitHub Action answers mentions of the Qwen coder bot to review or triage issues and pull requests, and twelve channels, among them Telegram, WeChat, DingTalk, Feishu, GitHub, GitLab and email, start tasks from messages, with persistent cron loops per channel target. Scheduled tasks inside a session fire on cron while the agent is running and expire after seven days.

Sourceqwenlm.github.io/qwen-code-docs/en/users/features/channels/overviewread 2026-11-08

Providers connect over OpenAI, Anthropic, Gemini or Vertex AI protocols, with DeepSeek, Grok, MiniMax, Z.AI, Kimi, OpenRouter and others built in, plus Qwen accounts, Alibaba Cloud plans and local models through Ollama, vLLM or LM Studio. A model command switches the main model and sets separate fast, vision and compaction models, and each subagent can run its own model. There is no automatic fallback when a request fails, apart from a vision fallback.

Sourceqwenlm.github.io/qwen-code-docs/en/users/configuration/model-providersread 2026-11-08

TypeScript, Python and Java SDKs drive the agent from code, the first marked experimental and the others alpha. Hooks on 23 events, from each tool call to session start and end, run commands, HTTP calls or prompts and can block a tool. The agent speaks MCP and the Agent Client Protocol, an experimental server exposes a REST API, headless runs return JSON, and channel plugins and extensions add new surfaces.

Sourceqwenlm.github.io/qwen-code-docs/en/users/features/hooksread 2026-11-08

The review command builds and tests changes in an isolated worktree, runs verifier agents and a reverse audit, and can post findings inline on a pull request. Goals use an independent verifier that checks whether the target was met, and the autofix command commits a pull request fix only when the repository's own checks pass. Arena compares several models on the same task in separate worktrees, though it is still experimental.

Sourceqwenlm.github.io/qwen-code-docs/en/users/features/code-reviewread 2026-11-08

Computer use runs through an MCP server or a native driver that exposes 19 desktop operations on macOS, Windows and Linux X11, once the operating system grants accessibility and screen recording rights. Browser use works through a Chrome extension that lets the agent operate the developer's signed in tabs on macOS or Linux, and can be switched off as a skill. Neither is marked experimental, though both need setup beyond the base install.

Sourceqwenlm.github.io/qwen-code-docs/en/users/features/computer-useread 2026-11-08