云启智联 MCP Server
OfficialREADME.md
# 云启智联 MCP Server
云启智联AI文档解析服务的 [Model Context Protocol (MCP)](https://modelcontextprotocol.io/) 服务器,让 Claude、Cursor、Windsurf 等支持 MCP 的 AI 客户端直接调用票据识别能力。
## 功能
- `parse_bank_receipt` — 银行回单解析(支持每页多张自动裁剪)
- `parse_bank_statement` — 银行对账单解析
- `parse_invoice` — 增值税发票解析
- `parse_document` — 通用文档解析
- `get_task_result` — 查询异步任务结果
## 前置要求
1. Python 3.10+
2. 云启智联 API Key(从 [http://8.135.62.13:5000](http://8.135.62.13:5000) 注册获取)
## 安装
```bash
pip install yqzl-mcp-server
```
或从源码安装:
```bash
git clone https://github.com/yunqi-zhilian/yqzl-mcp-server.git
cd yqzl-mcp-server
pip install -e .
```
## 启动
安装后,如果 `yqzl-mcp-server` 命令可用,直接运行:
```bash
yqzl-mcp-server
```
部分环境(尤其是 Windows)安装后脚本目录可能不在 PATH 中,导致命令找不到。此时可使用 Python 模块方式启动,效果完全相同:
```bash
python -m yqzl_mcp_server.server
```
## 在 Claude Desktop 中配置
编辑 `~/Library/Application Support/Claude/claude_desktop_config.json`(macOS)或 `%APPDATA%\Claude\claude_desktop_config.json`(Windows):
```json
{
"mcpServers": {
"yqzl-ai": {
"command": "yqzl-mcp-server",
"env": {
"YQZL_API_BASE_URL": "http://8.135.62.13:5000/AIService"
}
}
}
}
```
如果系统提示找不到 `yqzl-mcp-server` 命令(常见于 Windows),可将 `command` 改为使用 Python 模块方式启动:
```json
{
"mcpServers": {
"yqzl-ai": {
"command": "python",
"args": ["-m", "yqzl_mcp_server.server"],
"env": {
"YQZL_API_BASE_URL": "http://8.135.62.13:5000/AIService"
}
}
}
}
```
配置完成后重启 Claude Desktop,即可在对话中让 Claude 调用票据识别工具。
## 在 Cursor 中配置
打开 Cursor Settings → MCP,添加 server:
- Type: `stdio`
- Name: `yqzl-ai`
- Command: `yqzl-mcp-server`
如果命令找不到,可改用:
- Command: `python`
- Args: `-m yqzl_mcp_server.server`
## 使用示例
对 AI 说:
> 请用云启智联解析这张银行回单 `/Users/xxx/receipt.pdf`,API Key 是 `as-xxxxxxxx`。
AI 会自动调用 `parse_bank_receipt` 工具并返回结构化结果。
## 计费说明
MCP Server 不掌握计费逻辑。所有调用都通过你的云启智联 API Key 进行,按量从你在云启智联平台的余额中扣费。具体价格见官网。
## 自定义上游地址
默认使用 `http://8.135.62.13:5000/AIService`。域名备案完成后,可通过环境变量切换:
```bash
export YQZL_API_BASE_URL="https://yunqi-zhilian.com/AIService"
```
## 开发
```bash
pip install -e ".[dev]"
python -m yqzl_mcp_server.server
```
## 许可证
MIT
TDQS
A3.5/5.0
Scored across 5 tools
Disambiguation5/5
Each parse tool targets a distinct document type (bank receipt, bank statement, invoice, generic document), and get_task_result is clearly separate for async task management. No overlap.
Naming Consistency4/5
Most tools follow parse_* pattern, but get_task_result uses get_ instead. However, the action difference justifies the variation, and naming is otherwise consistent snake_case.
Tool Count5/5
5 tools is appropriate for a document parsing server, covering major document types and adding a task result retrieval tool without bloat.
Completeness4/5
Core document parsing types (receipt, statement, invoice, generic) are covered, and async result retrieval is included. Missing potentially other document types like contracts, but not a significant gap for the stated purpose.
Maintenance
ActivityInactive
ResponsivenessNo issues