101 lines
2.6 KiB
Markdown
101 lines
2.6 KiB
Markdown
|
# LiteLLM Proxy — Claude Code × xAI Grok
|
|||
|
|
|
|||
|
|
用 [LiteLLM](https://github.com/BerriAI/litellm) 在本地架一個 proxy server，讓 **Claude Code** 透過 Anthropic API 格式呼叫 **xAI Grok** 模型。
|
|||
|
|
|
|||
|
|
## 架構
|
|||
|
|
|
|||
|
|
```
|
|||
|
|
Claude Code
|
|||
|
|
│ ANTHROPIC_BASE_URL=http://localhost:4000
|
|||
|
|
▼
|
|||
|
|
LiteLLM Proxy (Docker, port 4000)
|
|||
|
|
│ model_list 路由
|
|||
|
|
▼
|
|||
|
|
xAI API (api.x.ai)
|
|||
|
|
└─ grok-4-1-fast-non-reasoning
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## 檔案結構
|
|||
|
|
|
|||
|
|
```
|
|||
|
|
litellm-proxy/
|
|||
|
|
├── docker-compose.yml # 啟動 LiteLLM Docker 容器
|
|||
|
|
├── litellm_config.yaml # 模型路由設定
|
|||
|
|
├── .env # xAI API Key（不要上傳到 git！）
|
|||
|
|
└── settings.json # 本地備份用的 claude settings 範例
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## 快速開始
|
|||
|
|
|
|||
|
|
### 1. 設定 API Key
|
|||
|
|
|
|||
|
|
編輯 `.env`：
|
|||
|
|
|
|||
|
|
```env
|
|||
|
|
XAI_API_KEY=xai-你的金鑰
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### 2. 啟動 Proxy
|
|||
|
|
|
|||
|
|
```bash
|
|||
|
|
docker compose up -d
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### 3. 設定 Claude Code
|
|||
|
|
|
|||
|
|
將以下內容寫入 `~/.claude/settings.json`：
|
|||
|
|
|
|||
|
|
```json
|
|||
|
|
{
|
|||
|
|
"env": {
|
|||
|
|
"ANTHROPIC_BASE_URL": "http://localhost:4000",
|
|||
|
|
"ANTHROPIC_AUTH_TOKEN": "sk-my-local-litellm-key",
|
|||
|
|
"ANTHROPIC_DEFAULT_SONNET_MODEL": "grok-4-1-fast-reasoning",
|
|||
|
|
"ANTHROPIC_DEFAULT_OPUS_MODEL": "grok-4-1-fast-reasoning",
|
|||
|
|
"ANTHROPIC_DEFAULT_HAIKU_MODEL": "grok-4-1-fast-reasoning"
|
|||
|
|
}
|
|||
|
|
}
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
> `ANTHROPIC_AUTH_TOKEN` 填的是 LiteLLM 的 `master_key`，不是真實的 Anthropic 或 xAI key。
|
|||
|
|
|
|||
|
|
### 4. 啟動 Claude Code
|
|||
|
|
|
|||
|
|
```bash
|
|||
|
|
claude
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## 常用指令
|
|||
|
|
|
|||
|
|
| 指令 | 說明 |
|
|||
|
|
|------|------|
|
|||
|
|
| `docker compose up -d` | 背景啟動 proxy |
|
|||
|
|
| `docker compose down` | 停止 proxy |
|
|||
|
|
| `docker compose logs -f` | 查看即時 log |
|
|||
|
|
| `docker compose down && docker compose up -d` | 重啟（套用新設定） |
|
|||
|
|
|
|||
|
|
## 設定說明
|
|||
|
|
|
|||
|
|
### `litellm_config.yaml`
|
|||
|
|
|
|||
|
|
```yaml
|
|||
|
|
model_list:
|
|||
|
|
- model_name: grok-4-1-fast-reasoning # Claude Code 看到的模型名稱
|
|||
|
|
litellm_params:
|
|||
|
|
model: xai/grok-4-1-fast-non-reasoning # 實際送給 xAI 的模型
|
|||
|
|
api_key: os.environ/XAI_API_KEY
|
|||
|
|
|
|||
|
|
litellm_settings:
|
|||
|
|
master_key: sk-my-local-litellm-key # Claude Code 用的假 API Key
|
|||
|
|
drop_params: true # 忽略 xAI 不支援的 Claude 專屬參數
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
**為什麼用 `non-reasoning` 模型？**
|
|||
|
|
Claude Code 本身是 agentic AI，reasoning 在 Claude Code 層進行。底層模型使用 non-reasoning 版本可避免 LiteLLM 的參數轉換 bug（`reasoningEffort` camelCase 問題），速度也更快。
|
|||
|
|
|
|||
|
|
## 注意事項
|
|||
|
|
|
|||
|
|
- `.env` 內含真實 API Key，請加入 `.gitignore` 避免上傳
|
|||
|
|
- 修改 `litellm_config.yaml` 後需重啟 docker 才會生效
|
|||
|
|
- `~/.claude/settings.json` 修改後**立即生效**，不需重啟
|