multi-model-router(多模型路由)

A Claude Code skill that lets one OpenAI-compatible API key power many AI tasks — automatically picking the right model for each job, asking you to confirm, then invoking it via Bash.

一个 Claude Code skill,让一把 OpenAI 兼容 API key 驱动多种 AI 任务——自动为每个任务挑选合适的模型,询问你确认后通过 Bash 调用。

English | 中文

---

English

Why?

You have one API key (e.g., from OpenRouter, SiliconFlow, DeepSeek, or your company's gateway) that exposes many models — Claude, GPT, DeepSeek, Kimi, Qwen, GLM, etc.

But you don't want to memorize "which model is best for which task". This skill makes Claude do it for you:

1. You describe the task ("transcribe this video", "convert this PDF to markdown", "summarize this 200K-word doc")
2. Claude identifies the task type, language, and constraints
3. Claude recommends a model with rationale ("for Chinese ASR, qwen3-asr-flash is the only top-tier option")
4. Claude asks you to confirm via an interactive prompt
5. Claude executes via Bash + curl, output back to you

Features

• 🔀 Smart routing — task → recommended model with reasoning
• 🔐 Secure setup — keys never enter conversation context, stored in ~/.config/api_keys.env (mode 600)
• 🌐 Provider-agnostic — works with any OpenAI-compatible gateway (OpenRouter, SiliconFlow, DeepSeek, Moonshot, custom)
• 🧠 Prompt source choice — uses your own prompt files (~/.config/llm-prompts/pdf-to-md.md, etc.) or built-in defaults; asks before overriding
• ⚡ One-command setup — bash scripts/setup.sh handles everything
• 📝 Bilingual docs — Chinese + English

Quick Start

# 1. Clone (or download as a Claude Code skill)
git clone https://github.com/YOUR_USERNAME/multi-model-router.git \
  ~/.claude/skills/multi-model-router

# 2. Run setup (you'll be prompted for provider + API key, hidden input)
bash ~/.claude/skills/multi-model-router/scripts/setup.sh

# 3. Verify (lists available models, runs a smoke test)
bash ~/.claude/skills/multi-model-router/scripts/verify.sh

# 4. In Claude Code, just describe a task — the skill activates automatically

Documentation

• SETUP.md — Full installation walkthrough
• MODELS.md — Task-to-model mapping reference
• FAQ.md — Common questions, including "Can I paste my API key into the chat?"
• examples/ — End-to-end examples

Security

🔒 Never paste your API key directly into a chat with Claude. Conversation history may be saved to local logs, knowledge bases, or third-party syncs. Always use setup.sh which reads keys via hidden input and stores them in a 600-permission file. See FAQ.md for details.

License

MIT. PRs welcome.

---

中文

为什么需要这个 skill?

你有一把 API key(可能来自 OpenRouter、硅基流动、DeepSeek 官方,或公司内部网关),它能调多种模型——Claude、GPT、DeepSeek、Kimi、Qwen、GLM 等等。

但你不想每次都记"哪个模型最适合哪类任务"。这个 skill 让 Claude 帮你判断:

1. 你描述任务("帮我转写这个视频"、"把这个 PDF 转成 markdown"、"总结这份 20 万字文档")
2. Claude 识别任务类型、语言、约束
3. Claude 推荐模型并说明理由("中文转写 qwen3-asr-flash 是唯一顶级选择")
4. Claude 询问确认(交互式提示)
5. Claude 执行 Bash + curl 调用,结果返回给你

特性

• 🔀 智能路由 — 任务 → 推荐模型,附理由
• 🔐 安全配置 — key 不进对话上下文,存于 ~/.config/api_keys.env(600 权限)
• 🌐 服务商无关 — 支持任意 OpenAI 兼容网关(OpenRouter / 硅基流动 / DeepSeek / Kimi / 自部署)
• 🧠 提示词复用 — 优先使用你自己的提示词文件(如 ~/.config/llm-prompts/pdf-to-md.md),没有才用默认
• ⚡ 一键安装 — bash scripts/setup.sh 全自动
• 📝 中英双语文档

快速开始

# 1. 克隆(或作为 Claude Code skill 下载)
git clone https://github.com/YOUR_USERNAME/multi-model-router.git \
  ~/.claude/skills/multi-model-router

# 2. 运行安装(会提示你选服务商 + 输入 API key,隐藏输入)
bash ~/.claude/skills/multi-model-router/scripts/setup.sh

# 3. 验证(列出可用模型,跑一次冒烟测试)
bash ~/.claude/skills/multi-model-router/scripts/verify.sh

# 4. 在 Claude Code 里直接描述任务,skill 会自动激活

文档

• SETUP.md — 完整安装指南
• MODELS.md — 任务-模型映射手册
• FAQ.md — 常见问题,包括"我能不能把 API key 直接粘到对话里?"
• examples/ — 端到端示例(PDF 转 md / 视频转写 / TTS 配音)

安全

🔒 永远不要把 API key 直接粘贴到与 Claude 的对话里。 对话历史可能被保存到本地日志、知识库、或第三方同步。请使用 setup.sh,它通过隐藏输入读取 key,存到 600 权限的文件里。详见 FAQ.md。

协议

MIT。欢迎 PR。