百度图片搜索 MCP Server
# 百度图片搜索 MCP Server
基于百度图片搜索 API 的 MCP (Model Context Protocol) 服务器。支持图片搜索、下载和格式转换一体化。
## 功能特性
- 🔍 **图片搜索**:搜索百度图片,返回图片标题、URL、尺寸、格式等信息
- 📥 **一键下载**:搜索同时可直接下载图片到本地
- 🔄 **格式转换**:支持 JPEG、PNG、WebP 三种格式互转(基于 sharp)
- 🖼️ **双源选择**:可选择缩略图(thumburl,百度 CDN 稳定快速)或原图(objurl,可能有防盗链)
- 📄 **分页支持**:支持自定义每页数量和页码
## 安装
```bash
# 克隆仓库
git clone <repo-url>
cd mcp_image_bd_tools
# 安装依赖
npm install
# 构建
npm run build
```
## 配置 MCP
在 MCP 客户端配置中添加:
```json
{
"mcpServers": {
"image-baidu": {
"command": "node",
"args": ["path/to/mcp_image_bd_tools/dist/index.js"]
}
}
}
```
## 使用
### 工具:`search_images`
#### 参数说明
| 参数 | 类型 | 必填 | 默认值 | 说明 |
|------|------|------|--------|------|
| `keyword` | string | ✅ | - | 搜索关键字 |
| `page` | number | ❌ | 1 | 页码,从 1 开始 |
| `perPage` | number | ❌ | 30 | 每页数量,最大 60 |
| `downloadPath` | string | ❌ | - | 下载目录路径。设置后自动下载图片 |
| `format` | string | ❌ | - | 输出格式:`jpeg` / `png` / `webp`(仅下载模式有效) |
| `urlType` | string | ❌ | `thumb` | 图片源:`thumb`(缩略图)/ `original`(原图) |
#### 使用示例
**仅搜索:**
```json
{
"keyword": "维多利亚港",
"page": 1,
"perPage": 10
}
```
**搜索并下载缩略图:**
```json
{
"keyword": "故宫",
"downloadPath": "./downloads",
"perPage": 5
}
```
**搜索并下载原图 + 转换为 JPEG:**
```json
{
"keyword": "黄山",
"downloadPath": "./downloads",
"format": "jpeg",
"urlType": "original",
"perPage": 3
}
```
#### 返回结果结构
```json
{
"keyword": "维多利亚港",
"totalNum": 44880,
"hasMore": true,
"page": 1,
"images": [
{
"index": 1,
"thumbUrl": "https://img0.baidu.com/it/u=...",
"originalUrl": "https://...",
"sourceUrl": "http://...",
"title": "图片标题",
"width": 1920,
"height": 1080,
"format": "jpg",
"isGif": false,
"isCopyright": false,
"detailUrl": "https://image.baidu.com/search/detail?...",
"downloaded": true,
"localPath": "downloads/1_图片标题_a1b2c3d4.jpg",
"convertedFrom": "webp→jpeg"
}
],
"error": "有 1 张图片下载失败(如:...)"
}
```
## 技术架构
```
src/
├── index.ts # MCP Server 入口,注册 search_images 工具
├── baidu-image.ts # 百度图片搜索 API 封装(acjson 接口)
├── downloader.ts # 图片下载 + sharp 格式转换模块
├── types.ts # TypeScript 类型定义
└── utils.ts # 工具函数(URL 构建、格式检测、随机 UA 等)
```
### 依赖
- **@modelcontextprotocol/sdk**: MCP 协议 SDK
- **axios**: HTTP 客户端
- **sharp**: 高性能图片格式转换(JPEG/PNG/WebP)
## 技术实现细节
### 搜索 API
使用百度图片搜索的 `acjson` 接口:
```
GET https://image.baidu.com/search/acjson
?tn=resultjson_com
&word={关键字}
&pn={偏移量}
&rn={每页数量}
&gsm={pn的十六进制}
&newReq=1
```
- **已验证:无需 Cookie 即可调用**,直接请求即可返回完整结果
- 分页通过 `pn`(偏移量,步长为 `rn`)和 `gsm`(`pn` 的十六进制值)控制
- 响应中包含 `totalNum`(总结果数)和 `hasMore`(是否有更多)
### 图片格式检测
下载图片时自动检测格式,优先级:
1. 响应头 `Content-Type`
2. 文件魔数(Magic Bytes)
如果设置了 `format` 参数且与实际格式不一致,自动使用 sharp 转换。
### 反爬策略
- 随机 User-Agent(从多个真实浏览器 UA 中随机选择)
- 正确的 Referer 头
- 请求间添加 300-700ms 随机延迟
- 并发下载数限制为 3
## License
MIT
TDQS
Scored across 1 tool
Only one tool exists, so there is zero risk of confusion between tools. The tool's internal modes (search, download, convert) are clearly described and mutually exclusive.
With a single tool named 'search_images', the naming is clean and follows a conventional verb_noun pattern. There are no inconsistent conventions to penalize.
Having only one tool is borderline for a typical MCP server, but the tool bundles search, download, and format conversion into one interface, which fits the server's stated purpose of Baidu image search. Still, it feels slightly thin compared to servers with dedicated tools for each operation.
The single tool covers the core lifecycle of image searching: querying with pagination, downloading thumbnails or originals, and converting formats. There are no obvious missing operations for a search-focused server.