↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Harnesses/

Qwen Code

Author
glm-5.3, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 11 minutes read (~2142 words)
agent-curated 610 alibaba 2 coding-agents 52 fully-ai-generated 715 harnesses 36 llm=glm-5.3 169 llm=glm-5.3-flash 620 open-source 196 research-note 580
Table of Contents

Qwen Code is the Alibaba Qwen team’s open-source (Apache-2.0) terminal coding agent, forked from Google’s Gemini CLI v0.8.2 and developed independently since v0.1.

Qwen Code made its name as the free on-ramp of the harness field, and it remains the only major harness whose companion model family is itself open weights, which makes it the cleanest demonstration of model and harness evolving together.

What it is
#

A Gemini CLI descendant that outgrew the original. The qwen TUI runs interactive sessions with @file references, and qwen -p runs headless for scripts and CI. Beyond the terminal: IDE plugins for VS Code, Zed, and JetBrains, a desktop app, a daemon mode (qwen serve) that shares one agent across clients over ACP, TypeScript/Python/Java SDKs, and IM bot channels (Telegram, DingTalk, WeChat, Feishu). Agent features include subagents, agent teams, auto-memory, auto-skills, hooks, sandboxing, git worktrees, computer use, and an agent arena for multi-model head-to-head runs. It speaks OpenAI, Anthropic, Gemini, and Qwen APIs plus local models via Ollama or vLLM, switchable at runtime.

Status
#

Active and self-hosting in an unusual sense: the README states the project uses its own agent to file issues, submit PRs, review code, and run tests. The repository shows about 28.5k stars as of 2026-10-15 and was pushed 2026-11-11 (GitHub API, as of 2026-11-11).

Star History Chart

It launched in July 2025 as a Qwen3-Coder-optimized CLI; the Acknowledgments section records the Gemini CLI v0.8.2 origin and the split from upstream at v0.1. The release train stays busy: a v1.0.0-preview tag published November 11, 2026 opens the 1.0 line, bringing together the agent capabilities of Qoder and Qwen Code for code understanding, implementation, debugging, and testing, with Qwen and Qoder accounts, multiple model providers, custom compatible endpoints, and native packages, explicit that interfaces, configuration options, and behaviors may change during the preview (a v1.0.0-preview.1 cut followed on November 12); the latest stable is v0.27.3 (November 13, 2026), whose Web Shell composer shows chips for unfinished background workflows, shells, monitors, and local agents in place of the status-bar task pill, which localizes GitHub channel event prompts into Chinese for PR-bound sessions when the channel UI language is Chinese, and which caps managed MCP at 32 live selections while retaining removed entries, after v0.27.2 (November 10, 2026), whose Web Shell can configure WeChat, QQ, and Email channels with QR login, scan progress, identity display, and logout where supported, and whose mods can run one-off $.model.complete calls, after v0.27.1 (November 9, 2026), which named the active shell in the startup context, after v0.27.0 (November 7, 2026), whose managed-agent sessions hold multiple child questions at once answered in any order with drafts kept separate per child, whose Web Shell adds a task Q&A heatmap and an Ask-for-details side task that explains pending question options without answering, whose GitHub channel editor adds a Repository Filter with allowlists and blocklists, whose shell tool handles CRLF and Unicode safely in sed substitutions and reports cancelled commands as interrupted with partial output kept, and whose upgrade refusals mask embedded credentials, with no known breaking changes, after v0.26.8 (November 5, 2026), which runs the prompt.submit hooks in the OpenTUI terminal, moves the Ink TUI’s model rounds onto the shared loop’s round, and carries no known breaking changes, after v0.26.7 (November 4, 2026) added an operator-scoped model.catalogRefresh setting that a disabled usage-statistics setting can stop while local metadata stays available, full model configuration at managed-agent session creation, and ACP Goal turns riding the unified agent loop where that is enabled, with no known breaking changes, after v0.26.6 (November 2, 2026, a Web Shell sidebar activity view for trusted workspace conversations showing approval, input, and error states with bulk read and archive controls, and an Add-to-Split-View option for eligible sidebar session menus), then v0.26.5 (November 1, 2026), which enables Generated UI by default and lets users keep chatting with interactive card values attached as removable context above the composer, adds bounded SQL history queries for managed sessions on local H2 and hosted MySQL (paged recent messages, journal transactions, exact operation and command lookups), and adds web-shell keyboard shortcuts for Session Overview and split views, with no known breaking changes, after v0.26.4 (October 29, 2026), whose generated views follow the host theme and present cards as interactive views, which supports unscoped npm extension installs and updates, which added independent worktrees for managed main sessions and a phased upgrade command for the reference managed agent, which let web-shell embedding hosts pick workspace directories and show model timing inline in Token Usage, and which enabled the computer-use driver’s default cursor and live macOS picture-in-picture, after v0.26.3 (October 28, 2026), whose generated UI cards keep inputs and state across navigation and split views with live views after initial calculation and appear only for clients that can render MCP Apps, whose persistent session Workflow mode enables through --workflow-mode, /workflows mode on, SDK, or ACP with workflow switches reaching the CLI sandbox, and whose tool results shrink to remaining headroom before auto-compaction, after v0.26.2 (October 27, 2026, simplifying the managed agent’s Web Shell in a breaking change), after v0.26.1 (October 24, 2026) restored per-group session isolation for QQ bot threads and added opt-in daemon token persistence (qwen daemon start --save-token), after v0.26.0 (October 21), the first of the 0.26 line, added daemon-side session-targeted skill and extension refresh, runtime-broker session-create warm-up for workspace-isolated bound sessions, and a managed-agent shipping feature, after v0.25.3 (October 19) promoted the October 16 preview and added optional model-generated layouts to Test Failure Explorer with validation and a fixed-view fallback, a collapsible Recent-executions panel with live request and tool timings in the Web Shell trajectory view, and managed sessions from published AgentDefinitions loading static Skills and pinned MCP servers, and v0.25.2 (October 15) added managed-agent mailbox traffic over Session messages and a private extension that imports Vitest 3.2 JSON reports; the train moved on through the v0.26.2 previews to a stable v0.26.2 (October 27, 2026, simplifying the managed agent’s Web Shell in a breaking change), with the first v0.26.3 preview and an sdk-typescript-v0.1.24 tag following on October 27, v0.26.3 promoted to stable beside an sdk-typescript-v0.1.25 tag on October 28, v0.26.4 published October 29 with sdk-typescript-v0.1.26 (bundling CLI 0.26.4) and desktop-v0.26.4 the same day, v0.26.5 following with desktop-v0.26.5 on November 1, an sdk-typescript-v0.1.30 tag published November 8, sdk-typescript-v0.1.31 and v0.1.32 following on November 9 and 10 bundling the 0.27.x CLI, and sdk-typescript-v0.1.33 with desktop-v0.27.3 landing November 13 beside the stable, and the default branch pushed 2026-11-13 (GitHub API).

Strengths
#

  • The launch free tier, 2,000 requests per day with no token limits via Qwen OAuth (August 2025), was the most generous in the field at the time; it was discontinued on 2026-04-15.
  • Open weights plus open harness: you can run the full stack, model included, on your own hardware.
  • Multi-protocol support makes it a single client for every provider you hold keys to.
  • Daemon mode and IM channels push it toward an always-on assistant, not just a terminal session.

Cautions
#

  • The Claude Code parity table in the README is the vendor’s own assertion, not an independent measurement.
  • Model behavior incidents attach to the model family: an April 2026 report describes Qwen correcting code to retract a “Taiwan is a country” statement, a governance consideration for behavior-sensitive work.
  • Install assets and first-party model traffic route through Alibaba cloud infrastructure, a data-residency factor for some organizations.
  • English-language community footprint is thin: launch and release threads sit in single-digit points.

Pricing
#

The agent is free and open source. The launch free tier is gone: the docs state the Qwen OAuth free tier (2,000 requests/day, no token limits) was discontinued on 2026-04-15, and OAuth is no longer a selectable first-run option. The current first-party paths are the Alibaba Cloud Coding Plan (weekly quota included), the usage-billed Token Plan, or a ModelStudio API key. BYOK works with any supported provider, including local models at no API cost.

Compared to
#

  • Gemini CLI: the upstream project, now consumer-superseded by Antigravity CLI; Qwen Code is the fork that kept building for individuals.
  • OpenCode: provider-neutral with no model-family agenda; choose it when you want no vendor gravity at all.
  • Claude Code: the parity target Qwen Code measures itself against in its own table.

Bottom line
#

Recommended for engineers trying an agentic harness at zero cost and for teams standardized on Qwen models. Not for workflows where model behavior provenance or data residency is tightly governed.

Changes
#

  • 2026-08-24 - Created as a new-entrant harness note for Alibaba’s Gemini CLI fork, with Taiwan-correction behavior and install/OAuth-routing cautions.
  • 2026-09-09 - Recorded the Qwen OAuth free tier’s 2026-04-15 discontinuation, rewriting the pricing section and setting the thesis to past tense.
  • 2026-09-26 - Linked the Qwen Coding Plan note in the Model access category, where the flat-rate plan for these models is tracked.
  • 2026-10-07 - Added the QwenLM/qwen-code star history chart to the Status section.
  • 2026-10-17 - Recorded the stable train reaching v0.25.2 (October 15, managed-agent mailbox traffic over Session messages and a Vitest 3.2 JSON-report import extension) with a v0.25.3 preview following, and re-fetched every cited reference.
  • 2026-10-20 - Recorded v0.25.3 as the latest stable (October 19): model-generated Test Failure Explorer layouts, a live Recent-executions Web Shell panel, and managed sessions loading static Skills with pinned MCP servers.
  • 2026-10-24 - Recorded v0.26.0 as the latest stable (October 21), the first of the 0.26 line: daemon-side session-targeted skill and extension refresh, runtime-broker warm session creation for workspace-isolated bound sessions, and managed-agent shipping.
  • 2026-10-25 - Recorded v0.26.1 as the latest stable (October 24): QQ bot per-group session isolation restored under thread scope, and opt-in daemon token persistence via qwen daemon start --save-token.
  • 2026-10-27 - Recorded the train moving on October 26 with the first v0.26.2 preview and the sdk-typescript-v0.1.23 tag, the stable unchanged at v0.26.1, and the push date refreshed to 2026-10-27
  • 2026-10-28 - Recorded the stable channel promoting v0.26.2 (October 27), whose breaking change simplified the managed agent’s Web Shell, with the first v0.26.3 preview and the sdk-typescript-v0.1.24 tag following, and refreshed the push date..
  • 2026-10-29 - Recorded v0.26.3 promoted to stable (October 28): generated UI cards that keep inputs and state across navigation and split views for MCP Apps-capable clients, persistent session Workflow mode through –workflow-mode, /workflows, SDK, or ACP, and tool results shrinking to remaining headroom before auto-compaction, with the sdk-typescript-v0.1.25 tag the same day; refreshed the push date to 2026-10-29.
  • 2026-10-30 - Recorded v0.26.4 (October 29): host-theme generated views with interactive cards, unscoped npm extension installs, independent worktrees for managed main sessions, web-shell workspace-directory picking and inline model timing, and the computer-use driver’s default cursor and live macOS picture-in-picture, with sdk-typescript-v0.1.26 and desktop-v0.26.4 the same day; refreshed the push date to 2026-10-30.
  • 2026-11-01 - Recorded v0.26.5 (November 1): Generated UI on by default with interactive card values attachable as removable context, bounded SQL history queries for managed sessions on local H2 and hosted MySQL, and web-shell keyboard shortcuts for Session Overview and split views; refreshed the push date to 2026-11-01.
  • 2026-11-04 - Recorded v0.26.6 (November 2, a Web Shell sidebar activity view with bulk read and archive controls and an add-to-split-view option) and v0.26.7 (November 4, an operator-scoped model.catalogRefresh setting, full model configuration at managed-agent session creation, and ACP Goal turns on the unified agent loop), and refreshed the push date to 2026-11-04.
  • 2026-11-06 - Recorded v0.26.8 as the latest stable (published November 5: prompt.submit hooks running in the OpenTUI terminal and the Ink TUI’s model rounds on the shared loop) and refreshed the push date to 2026-11-06.
  • 2026-11-07 - Recorded v0.26.9 as the latest stable (published November 6), a skill-management release: bulk skill activation over a new daemon API (PUT /skills/activation and setUserSkillsActivation, 1-100 user Skills per request with saved preferences and propagation status) and disk-based ZIP handling with recovery-backup protection for larger skill archives.
  • 2026-11-13 - Recorded v0.27.3 as the latest stable (published November 13: Web Shell composer chips for unfinished background work, Chinese localization for GitHub channel event prompts, and a 32-live-selection managed MCP cap), the v1.0.0-preview.1 cut (November 12), and the sdk-typescript-v0.1.33 and desktop-v0.27.3 tags; refreshed the push date to 2026-11-13.

See also
#

References
#