Route tasks from Claude Code to 7 Chinese AI model backends. Each backend runs as an isolated Claude Code instance with its own API provider and provider-specific profile controls.
Inspired by openai/codex-plugin-cc — the plugin architecture that made multi-model delegation in Claude Code possible. Thank you, OpenAI.
| Model | Backend (default tier) | Strength | Command |
|---|---|---|---|
| Doubao | doubao-seed-code-preview-latest | General Chinese coding, frontend/vision coding, Seed 2.0/router fallback | /cn:doubao |
| Qwen | qwen3-coder-next (opus → qwen3-coder-plus) | Agentic coding, SQL / Alibaba ecosystem, Token Plan/PayG routes | /cn:qwen |
| Kimi | kimi-for-coding | Stable Kimi Code route, long context, 64K out | /cn:kimi |
| GLM | glm-4.7 (opus → glm-5.1) | Reasoning / Chinese understanding, Z.ai Claude Code route | /cn:glm |
| StepFun | step-3.5-flash-2603 (sonnet+) | Math / logic, vision, 64K out | /cn:stepfun |
| MiniMax | MiniMax-M2.7 | Stable M2 route, highspeed/cheap profiles, 64K Anthropic out | /cn:minimax |
| MiMo | mimo-v2-pro (Token Plan SGP) | Xiaomi flagship, 1M context, V2.5/Omni/Flash profiles | /cn:mimo |
/plugin marketplace add LeoLin990405/cn-cc
/plugin install cn@cn-cc
/reload-plugins
/cn:setupOr add to ~/.claude/settings.json:
{
"enabledPlugins": {
"cn@cn-cc": true
},
"extraKnownMarketplaces": {
"cn-cc": {
"source": {
"source": "github",
"repo": "LeoLin990405/cn-cc"
}
}
}
}/cn:setupCN Models Setup — 7/7 available
✓ doubao Doubao (doubao-seed-code-preview-latest) 2.1.150 (Claude Code)
✓ qwen Qwen (qwen3-coder-next; opus→qwen3-coder-plus) 2.1.150 (Claude Code)
✓ kimi Kimi (kimi-for-coding) 2.1.150 (Claude Code)
✓ glm GLM (glm-4.7; opus→glm-5.1) 2.1.150 (Claude Code)
✓ stepfun StepFun (step-3.5-flash-2603) 2.1.150 (Claude Code)
✓ minimax MiniMax (MiniMax-M2.7) 2.1.150 (Claude Code)
✓ mimo MiMo (mimo-v2-pro) 2.1.150 (Claude Code)
Direct commands pass arguments through to cn-companion.mjs, so provider profiles can be selected inline:
/cn:qwen --profile token 帮我做一次复杂代码审查
/cn:glm --profile max 分析这个性能瓶颈
/cn:doubao --profile vision 检查这个前端组件
/cn:minimax --profile highspeed 快速总结这批日志
/cn:mimo --profile latest 处理一个长上下文多模态任务If the prompt itself starts with flags, separate command options from prompt text with --:
/cn:qwen --profile token -- --json 这个参数是什么意思？Use /cn:setup --doctor for the deeper wrapper health check, or run the companion directly:
node "${CLAUDE_PLUGIN_ROOT}/scripts/cn-companion.mjs" profiles/cn:ask 帮我写一个 Doris 数据仓库的 ETL SQL # → Qwen
/cn:ask 分析这篇 8 万字的研究报告 # → Kimi
/cn:ask 证明这个不等式 # → StepFun
/cn:ask 写一个 Python 爬虫 # → DoubaoThe cn-dispatch agent reads task signals and picks the best model:
| Signal | Routes to | Why |
|---|---|---|
| SQL / Doris / ADB / PolarDB | Qwen | Alibaba ecosystem native |
| Long text 50K–200K tokens | Kimi | stable Kimi Code long-context route |
| Ultra-long context / multimodal | MiMo | token-plan Pro plus V2.5/Omni profiles |
| Math / proofs / logic | StepFun | Math specialist |
| Deep reasoning / Chinese NLU | GLM | Strong Chinese reasoning |
| Quick / lightweight tasks | MiniMax | Stable/highspeed M2 profiles |
| General Chinese coding | Doubao | Best all-round (default) |
/cn:kimi <prompt> # Long context
/cn:qwen <prompt> # SQL / Alibaba
/cn:glm <prompt> # Reasoning
/cn:doubao <prompt> # General coding
/cn:stepfun <prompt> # Math / logic
/cn:minimax <prompt> # High-speed
/cn:mimo <prompt> # Xiaomi MiMo, 1M ctxThe cn-dispatch agent is also triggered automatically by Claude when it detects a task that would benefit from a Chinese model backend. No slash command needed.
Claude Code (main session, Claude Opus/Sonnet)
│
├─ /cn:ask "prompt" ← user-triggered smart routing
│ └─ cn-dispatch agent
│ ├─ cn-routing skill → selects model
│ └─ cn-companion.mjs task --model <name> "prompt"
│ └─ cc-<name> -p "prompt" --max-turns 1
│ └─ isolated CC instance → provider API
│
├─ /cn:kimi "prompt" ← user-triggered direct
│ └─ cn-companion.mjs task --model kimi "prompt"
│
└─ cn-dispatch agent ← auto-triggered by Claude
└─ (same flow as /cn:ask)
Each model needs a wrapper script in ~/bin/:
#!/usr/bin/env bash
# ~/bin/cc-<provider>
REAL_HOME="$HOME"
export HOME="$REAL_HOME/.claude-envs/<provider>"
mkdir -p "$HOME"
export ANTHROPIC_BASE_URL="<provider-api-endpoint>"
export ANTHROPIC_AUTH_TOKEN="$<PROVIDER_API_KEY>"
export ANTHROPIC_MODEL="<model-id>"
export API_TIMEOUT_MS="3000000"
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC="1"
exec claude "$@"Set these environment variables (e.g. in ~/.zshrc):
export ARK_API_KEY="..." # Doubao (Volcengine)
export DASHSCOPE_API_KEY="..." # Qwen (Alibaba)
export KIMI_API_KEY="..." # Kimi (Moonshot)
export GLM_API_KEY="..." # GLM (Zhipu)
export STEPFUN_API_KEY="..." # StepFun
export MINIMAX_API_KEY="..." # MiniMax
export MIMO_API_KEY="tp-..." # MiMo Token Plan (SGP region, key starts with tp-)plugins/cn/
├── agents/
│ └── cn-dispatch.md # Smart routing agent
├── commands/
│ ├── setup.md # /cn:setup
│ ├── ask.md # /cn:ask (smart routing)
│ ├── status.md # /cn:status
│ ├── doubao.md # /cn:doubao
│ ├── qwen.md # /cn:qwen
│ ├── kimi.md # /cn:kimi
│ ├── glm.md # /cn:glm
│ ├── stepfun.md # /cn:stepfun
│ ├── minimax.md # /cn:minimax
│ ├── mimo.md # /cn:mimo
│ └── profiles.md # /cn:profiles
├── skills/
│ ├── cn-routing/SKILL.md # Model selection decision matrix
│ └── cn-result-handling/SKILL.md # Output formatting rules
└── scripts/
├── cn-companion.mjs # Core runtime
└── cn-companion.test.mjs # CLI parser smoke tests
launchers/
├── bin/ # Snapshot of the local cc-* wrappers
└── prompts/ # Provider-specific appended system prompts
This project would not exist without the pioneering work of the Codex plugin for Claude Code by OpenAI. Their plugin architecture — commands, agents, skills, and companion scripts — provided the blueprint that made multi-model delegation in Claude Code practical. We are grateful for their contribution to the open-source ecosystem.