Skip to main content
The Qwen Code converter creates YAML agent files, Markdown commands, and extracts settings from MCP environment variables.

​
Installation

bunx @every-env/compound-plugin install compound-engineering --to qwen

​
Output Structure

Writes to ~/.qwen/extensions/<plugin-name>/: 
~/.qwen/extensions/compound-engineering/
├── qwen-extension.json       # Extension config
├── QWEN.md                   # Context file
├── agents/
│   └── *.yaml                # Agent configs
├── commands/
│   ├── simple.md             # Flat commands
│   └── workflows/
│       └── plan.md           # Nested commands (workflows:plan)
└── skills/
    └── */                    # Pass-through skills
Qwen uses YAML for agents and Markdown for commands. Nested commands use directory structure.

​
Conversion Details

​
Extension Config

qwen-extension.json: 
{
  "name": "compound-engineering",
  "version": "1.0.0",
  "commands": "commands",
  "skills": "skills",
  "agents": "agents",
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem"],
      "env": {
        "ALLOWED_PATHS": "/Users/you/projects"
      }
    }
  },
  "settings": [
    {
      "name": "Filesystem Access Token",
      "description": "Environment variable for filesystem MCP server",
      "envVar": "FILESYSTEM_TOKEN",
      "sensitive": true
    }
  ]
}
The settings array is auto-extracted from MCP server env vars that look like placeholders (${VAR}, YOUR_KEY, etc.).

​
Context File

QWEN.md (auto-generated): 
# compound-engineering

Compound Engineering Plugin

## Agents

- **plan-specialist**: Planning agent for complex features
- **code-review-agent**: Review code for issues

## Commands

- **/workflows:plan**: Turn feature ideas into detailed plans
- **/ce:work**: Execute plans with worktrees and task tracking

## Skills

- compound-engineering-skill
- repo-research
Provides overview of plugin capabilities.

​
Agents → YAML Files

Agents are converted to YAML: agents/plan-specialist.yaml: 
---
name: plan-specialist
description: Planning agent for complex features
model: anthropic/claude-sonnet-4-5
temperature: 0.2
---

You are a planning specialist...
Frontmatter fields:
  • name: Agent name
  • description: Agent description
  • model: Normalized model with provider prefix
  • temperature: Inferred from agent name/description

​
Commands → Markdown Files

Commands become .md files: commands/workflows/plan.md (for workflows:plan): 
---
description: Turn feature ideas into detailed plans
model: anthropic/claude-sonnet-4-5
allowedTools:
  - Read
  - Write
---

You are a planning specialist...
Nested structure: 
workflows:plan → commands/workflows/plan.md
ce:work → commands/ce/work.md
simple → commands/simple.md

​
Settings Extraction

Environment variables with placeholders are extracted as settings: Input (MCP server env): 
{
  "env": {
    "FILESYSTEM_TOKEN": "${FILESYSTEM_TOKEN}",
    "API_KEY": "YOUR_API_KEY",
    "ALLOWED_PATHS": "/Users/you/projects"
  }
}
Output (in qwen-extension.json): 
{
  "settings": [
    {
      "name": "Filesystem Token",
      "description": "Environment variable for filesystem MCP server",
      "envVar": "FILESYSTEM_TOKEN",
      "sensitive": true
    },
    {
      "name": "Api Key",
      "description": "Environment variable for filesystem MCP server",
      "envVar": "API_KEY",
      "sensitive": true
    }
  ]
}
Sensitive detection: A setting is marked sensitive: true if the env var name contains:
  • KEY
  • TOKEN
  • SECRET
Only env vars that look like placeholders are extracted. Concrete values like /Users/you/projects are not.

​
Path Rewriting

Claude paths are rewritten: 
- ~/.claude/skills/
+ ~/.qwen/skills/
- .claude/
+ .qwen/

​
MCP Servers

Only stdio servers are supported: 
{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem"],
      "env": {
        "ALLOWED_PATHS": "/Users/you/projects"
      }
    }
  }
}
Remote MCP servers (HTTP/SSE) are not supported and will be skipped with a warning.

​
Model Normalization

Bare model aliases are converted to provider-prefixed names:
InputOutput
haikuanthropic/claude-haiku
sonnetanthropic/claude-sonnet
opusanthropic/claude-opus
claude-3-5-sonnetanthropic/claude-3-5-sonnet
gpt-4openai/gpt-4
gemini-progoogle/gemini-pro
qwen-maxqwen/qwen-max

​
Temperature Inference

When inferTemperature: true, the converter analyzes agent names:
PatternTemperature
review, audit, security0.1
plan, architecture0.2
doc, readme, changelog0.3
brainstorm, creative0.6
Default(none)

​
Example Output

Input (Claude agent): 
{
  "name": "plan-specialist",
  "description": "Planning agent",
  "model": "sonnet",
  "body": "You are a planning specialist."
}
Output (~/.qwen/extensions/compound-engineering/agents/plan-specialist.yaml): 
---
name: plan-specialist
description: Planning agent
model: anthropic/claude-sonnet
temperature: 0.2
---

You are a planning specialist.
Input (Claude command): 
{
  "name": "workflows:plan",
  "description": "Create plan",
  "body": "You plan features."
}
Output (~/.qwen/extensions/compound-engineering/commands/workflows/plan.md): 
---
description: Create plan
---

You plan features.

​
Limitations

  • stdio MCP only: Remote MCP servers are not supported
  • No hooks support: Qwen does not have a hooks equivalent
  • Settings extraction only: Settings are extracted from env vars, not user-configurable
  • Nested command structure: Commands with colons create nested directories
If your plugin has hooks, they will not be converted. Qwen does not support lifecycle hooks.

​
Sync Support

Sync personal Claude config to Qwen: 
bunx @every-env/compound-plugin sync --target qwen
This syncs:
  • Personal skills from ~/.claude/skills/ (symlinked)
  • Personal commands from ~/.claude/commands/ (as Qwen commands)
  • MCP servers from ~/.claude/settings.json (stdio only)

​
See Also

Qwen Code Documentation

Official Qwen Code documentation

Sync Command

Sync personal Claude config to Qwen