Alibaba’s Qoder 1.0 Transforms Desktop AI Coding – Hands‑On Review

Core Changes in Qoder 1.0

Qoder 1.0 redesigns the product from a prototype to a full‑desktop AI workbench. Five main themes are introduced:

Multi‑Agent Parallelism – multiple agents can work concurrently instead of a single AI writing code.

End‑to‑End Delivery – a complete pipeline Plan → Code → Verify replaces the previous “write‑and‑hand‑off” flow.

Long‑Term Memory – the system accumulates project conventions and technology‑stack knowledge the longer it is used.

Extensible Expert Team – built‑in experts can be supplemented with custom ones to inject industry knowledge or team standards.

Full‑Link Quality Check – AI performs code review, verification, and computer‑use checks before delivering results.

Quest View – New Primary UI

Quest, formerly a hidden chat mode, becomes an independent three‑column view:

Left Pane – navigation, task list, workspace switching, knowledge and agent management.

Middle Pane – conversation flow with aggressive folding; results appear prominently.

Right Pane – output area with three tabs:

Spec – view or download specification documents.

Changed Files – accept or discard generated code changes.

Preview – live preview of execution results.

Each Quest shows a status (Running, Action Required, Ready, Error) so tasks can be left unattended and resumed later.

Model Selector

The selector offers two selection modes:

Tiered Choice – Auto (default routing, ~1.0× credit) for everyday development and Ultimate for complex system design (no credit multiplier shown).

Specific Model List – Qwen3.6‑Plus, DeepSeek‑V4‑Pro, DeepSeek‑V4‑Flash, GLM‑5.1, Kimi‑K2.6, MiniMax‑M2.7, with credit multipliers as low as 0.1×.

Users can also provide their own API keys (BYOK) to use custom models.

Agent Mode & Expect Mode

After the Quest upgrade, AI operates as a customizable engineering team rather than a single agent. A Team Lead interprets goals, breaks them into subtasks, coordinates agents, and enforces quality. Expert agents run in parallel without blocking each other. Internal tests report a quality improvement of roughly 67 % compared with a single‑agent setup.

Custom Business Experts

When built‑in experts are insufficient, users can define custom experts composed of:

SKILL – task‑specific methods such as requirement analysis, interface design, security checks, test generation, or code review.

MCP – connectors that bring existing tools, knowledge bases, or R&D systems into the execution context.

Creation methods:

Interactive command: /create-skill <description> Skills CLI installation, e.g.: npx skills add vercel-labs/agent-browser -a qoder Manual placement of SKILL.md files under ~/.qoder/skills/ or .qoder/skills/, after which the / command list reflects the new skill.

Skills Extension Mechanism

Each Skill is a SKILL.md file that describes when the model should invoke it. Three acquisition paths exist:

Built‑in assistant creation via /create-skill.

CLI installation from the Skills marketplace, e.g., npx skills add vercel-labs/agent-browser -a qoder.

Manual file placement as described above.

Built‑in high‑frequency commands include: /create-skill-ui – generate an interactive HTML widget for a Skill. /vercel-deploy – one‑click deployment to Vercel. /create-subagent – create a custom sub‑intelligent‑agent. /generate-structured-prd – produce a structured PRD. /init – scan the repository and generate AGENTS.md. /review – analyze uncommitted Git changes and suggest improvements.

Repo Wiki – Living Code Encyclopedia

Repo Wiki automatically generates structured documentation for a project and tracks code changes. When the code diverges, the Wiki flags inconsistencies; a single click updates the docs. The generated Wiki resides in .qoder/repowiki, can be committed to a remote branch, and teammates can retrieve it with git pull. It supports Chinese and English subfolders ( repowiki/zh/, repowiki/en/) and is limited to 10 000 files per project (Git repository with at least one commit required).

Browser Agent

Quest includes a browser agent capable of opening pages, clicking buttons, filling forms, scrolling, and taking screenshots. Two execution options are provided:

Built‑in lightweight panel – quick preview within the IDE.

Local Chrome – full‑featured Chrome instance for complex web apps.

The agent can be triggered automatically or explicitly with the /browser command.

Context System

The @ reference system now supports four types: @file – reference a code or rule file. @folder – reference an entire directory. @attachments – attach any local file (md, xmind, xlsx, docx, pdf, jpg, png) for the model to consume. @rule – inject project rules into the model’s system prompt.

This enables workflows such as dragging a design mockup to generate front‑end code, importing an XMind file to understand architecture, or attaching a spreadsheet for data‑structure analysis.

Execution Environments

When creating a Quest, users select an execution environment:

Local – modifies the main workspace directly; zero startup cost; suited for simple tasks and quick validation.

Worktree – runs in a hidden worktree, keeping the main branch clean; suited for medium‑complexity tasks and parallel multi‑task work; supports easy rollback.

Hands‑On Demo: Recreating a Browser Extension

A practical test cloned a social‑media anti‑addiction Chrome extension. The workflow was:

Provide a vague prompt containing the original extension’s Chrome URL (e.g., chrome://extensions/?id=elbikiflgfhjdjmficnigpeegjbhdidh).

The browser agent scraped the original UI, generated a TODO list, and wrote the code.

The system performed an automatic self‑review.

Additional minor requests were added with a follow‑up prompt and applied instantly.

The resulting extension was installed via Chrome; the settings page allowed arbitrary domain addition and activation‑duration adjustment.

The entire process followed a “task → sip coffee → view summary → apply” rhythm, illustrating the “Quest On, Hands Off” experience.

Summary of Qoder 1.0

Four keywords capture the upgrade:

Form – Quest upgraded to an independent three‑column view, eliminating fragmented AI‑as‑a‑plugin experiences.

Parallel – multiple workspaces, quests, and expert agents run concurrently.

Retention – the knowledge engine plus Repo Wiki continuously learn project and team specifics.

Domestic‑Friendly – native Chinese support, domestic model lineup, and direct network connectivity.