Clinical Handover MCP Server
Provides structured tools for risk classification, SBAR generation, completeness validation, follow-up task extraction, and ward preference management for clinical handover coordination agents.
README
Clinical Handover MCP Server
A Model Context Protocol (MCP) server for clinical handover coordination agents. It provides structured tools for risk classification, SBAR generation, completeness validation, follow-up task extraction, and ward preference management.
⚠️ Safety Notice: All outputs are clinical communication support drafts only. This system does not diagnose, prescribe, or make clinical decisions. Every output must be reviewed by a qualified clinical professional before use in patient care.
Quick Start
Prerequisites
- Node.js ≥ 20
- npm ≥ 9
Install & Build
# 1. Install dependencies
npm install
# 2. Compile TypeScript → build/
npm run build
# 3. Verify the server starts cleanly
npm start
# Expected stderr: [clinical-handover-mcp] Server running on stdio transport. Ready.
# Press Ctrl+C to stop.
Local Development (no build step)
npm run dev # runs src/index.ts via tsx directly
Lint (TypeScript type-check without emitting)
npm run lint
Inspect with MCP Inspector
npm run inspect
# Opens MCP Inspector UI — usually at http://localhost:5173
The inspector lets you call each tool interactively with a form UI and see raw
JSON responses. Use the sample inputs in src/data/sample-handover.md as test data.
Tools
| Tool | Purpose |
|---|---|
classify_patient_risk |
Risk-score a patient case (high/medium/low/uncertain) |
validate_handover_completeness |
Check for missing critical fields, return 0–100 score |
generate_follow_up_tasks |
Extract prioritised tasks from Gmail/Fireflies/Notion text |
build_sbar_handover |
Format a structured SBAR handover document |
create_handover_record |
Assemble a full shift handover record |
update_ward_preferences |
Draft ward-specific preference rules from clinician feedback |
Connect to Claude Desktop
Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
or %APPDATA%\Claude\claude_desktop_config.json (Windows):
{
"mcpServers": {
"clinical-handover": {
"command": "node",
"args": ["/ABSOLUTE/PATH/TO/clinical-handover-mcp-server/build/index.js"]
}
}
}
Replace /ABSOLUTE/PATH/TO/ with the real path, then restart Claude Desktop.
The 6 tools will appear in Claude's tool palette.
Connect to Agentman
In your Agentman agent config, add this MCP server as a tool source:
{
"mcp_servers": [
{
"name": "clinical-handover",
"transport": "stdio",
"command": "node",
"args": ["build/index.js"],
"cwd": "/path/to/clinical-handover-mcp-server"
}
]
}
The agent will discover all 6 tools automatically via the MCP protocol.
Environment Variables
Copy .env.example to .env and populate if needed:
cp .env.example .env
Currently the server needs no secrets — all logic is local. Future integrations (e.g. Notion write-back) would add tokens here.
Future: Streamable HTTP Deployment
The server currently uses stdio transport (simplest for local agents and Claude Desktop).
To expose it as an HTTP endpoint for multi-agent or remote deployments:
-
Install the HTTP transport package when available:
npm install @modelcontextprotocol/sdk-transport-http -
Replace
StdioServerTransportinsrc/index.tswithStreamableHttpServerTransport:import { StreamableHttpServerTransport } from "@modelcontextprotocol/sdk-transport-http"; const transport = new StreamableHttpServerTransport({ port: 3000 }); -
Deploy behind a reverse proxy (nginx/Caddy) with TLS.
-
Add bearer token auth middleware before exposing publicly.
For now, stdio is preferred — it keeps the attack surface minimal and avoids network credential management for a clinical communication tool.
Safety Design Principles
- No diagnosis. Tools score and classify for communication purposes only — not to inform clinical treatment.
- No prescribing. The
recommendationfield in SBAR is a handover communication field, not a prescription. - Mandatory disclaimer. Every tool output carries the safety notice.
- Human approval gate.
update_ward_preferencesnever writes to Notion directly — it drafts rules for human review. - No PII storage. The server holds no state between calls. Patient identifiers used in tool calls are not persisted.
Project Structure
src/
index.ts Entry point — stdio transport setup, graceful shutdown
server.ts Tool registration (MCP tool schemas + handlers)
logic.ts Core business logic (risk scoring, completeness, task extraction)
formatters.ts Output formatters (SBAR markdown, task tables, handover records)
safety.ts Safety disclaimer constants
types.ts Shared TypeScript interfaces and type aliases
data/
sample-handover.md Sample handover text for testing
docs/
CODEX_PROMPT.md Agent system prompt reference
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。