KnowS MCP Server
# KnowS MCP Server
KnowS 医学证据检索与分析平台的 MCP (Model Context Protocol) Server。
## 安装与使用
### 前置条件
本包发布在 GitHub Packages,需要先配置 npm registry。在 `~/.npmrc` 中添加:
```
//npm.pkg.github.com/:_authToken=YOUR_GITHUB_TOKEN
@PancrePal-xiaoyibao:registry=https://npm.pkg.github.com
```
> `YOUR_GITHUB_TOKEN` 需要有 `read:packages` 权限。
### 环境变量
| 变量 | 必填 | 说明 |
|------|------|------|
| `KNOWS_API_KEY` | 是 | KnowS API 密钥 |
| `KNOWS_API_HOST` | 否 | API 地址,默认 `https://dev-api.nullht.com` |
### Claude Code 配置
在 `~/.claude/settings.json` 或项目 `.claude/settings.json` 中添加:
```json
{
"mcpServers": {
"knows": {
"command": "npx",
"args": ["@PancrePal-xiaoyibao/knows-mcp-server"],
"env": {
"KNOWS_API_KEY": "your_api_key",
"KNOWS_API_HOST": "https://api.nullht.com"
}
}
}
}
```
### Claude Desktop 配置
在 `~/Library/Application Support/Claude/claude_desktop_config.json`(macOS)中添加:
```json
{
"mcpServers": {
"knows": {
"command": "npx",
"args": ["@PancrePal-xiaoyibao/knows-mcp-server"],
"env": {
"KNOWS_API_KEY": "your_api_key",
"KNOWS_API_HOST": "https://api.nullht.com"
}
}
}
}
```
### 直接运行
```bash
# 通过 npx
KNOWS_API_KEY=your_key npx @PancrePal-xiaoyibao/knows-mcp-server
# 或全局安装后运行
npm install -g @PancrePal-xiaoyibao/knows-mcp-server
KNOWS_API_KEY=your_key knows-mcp-server
```
## 本地开发
```bash
# 安装依赖
npm install
# 使用 bun 开发(热重载)
KNOWS_API_KEY=your_key bun run dev
# 构建
npm run build
# 运行构建产物
KNOWS_API_KEY=your_key npm start
```
## 工具列表
| 工具 | 说明 |
|------|------|
| `ai_search` | AI 检索,根据问题返回相关医学证据列表 |
| `evidence_summary` | 单篇证据 AI 概要总结 |
| `all_evidence_summary` | 获取所有证据的 AI 总结 |
| `evidence_highlight` | 获取证据被引用的原文内容 |
| `answer` | 基于问题和证据生成场景总结 |
| `answer_stream` | 场景总结(流式收集) |
| `auto_tagging` | 证据自动化标签提取 |
| `get_paper_en` | 英文文献详情 |
| `get_paper_cn` | 中文文献详情 |
| `get_guide` | 指南详情 |
| `get_meeting` | 会议详情 |
| `list_questions` | 历史提问列表 |
| `list_interpretations` | 文献解读历史列表 |
| `create_evidence_by_pdf` | 通过 PDF 创建新证据 |
## 系统提示词
项目附带了推荐的系统提示词 [`SYSTEM_PROMPT.md`](./SYSTEM_PROMPT.md),用于指导 AI 助手基于 KnowS 工具进行结构化的医学文献调研。
## 发版
```bash
# 更新 package.json 中的 version 后
git tag v0.x.x
git push origin v0.x.x
```
推送 `v*` tag 后,GitHub Actions 会自动运行测试并发布到 GitHub Packages。
TDQS
Scored across 14 tools
Most tools have distinct purposes (e.g., ai_search vs. answer vs. evidence_summary). However, answer and answer_stream are clearly variants of the same operation, and all_evidence_summary and evidence_summary overlap in scope, causing slight ambiguity.
Naming patterns are mixed: some tools use verb_noun (get_guide, list_questions) while others use noun_verb or noun_noun (evidence_highlight, ai_search). The presence of answer and answer_stream as single verbs breaks the pattern, reducing consistency.
With 14 tools covering search, summarization, evidence management, and history, the count feels well-scoped for a medical knowledge server. Each tool serves a clear role without unnecessary redundancy.
The tool set covers core workflows like searching, answering, and summarizing, but lacks update/delete operations for evidence and does not provide listing tools for guides or meetings, leaving notable gaps in lifecycle coverage.