Playbook: Patient Requests via MCP

MCP (Model Context Protocol) is the "USB-C for AI apps"—a standard that lets assistants like Claude talk to small "servers" that expose tools and data you choose.

Claude Desktop (and others) can connect to your local or hosted MCP server with a tiny JSON config; the server declares tools like create_ticket, search_FAQ, appt_slots.

Start with low-risk, high-volume requests:

• Medication refill intake (collect required fields, sanity checks, create a task)
• Appointment options (read calendar slots and propose times)
• Records request (collect identity + destination, generate a task)
• FAQs (search a clinic FAQ doc and draft answers for staff approval)

• Claude Desktop installed (Mac/Windows)
• Node.js 18+ (for the TypeScript server)
• A Google Sheet or lightweight DB (even a JSON file) to store request tickets
• Optional: A starter MCP server template (TypeScript)

A. Create a folder

mkdir mcp-patient-requests && cd mcp-patient-requests
npm init -y
npm install @modelcontextprotocol/sdk zod
npm install -D typescript ts-node @types/node
npx tsc --init

B. Add src/server.ts — two tools to start:

• create_request → validates and writes a ticket to data/requests.json
• list_requests → returns open tickets for follow-up

import { Server, Tool } from "@modelcontextprotocol/sdk/server";
import { z } from "zod";
import { readFileSync, writeFileSync, existsSync, mkdirSync } from "fs";

const server = new Server({ name: "patient-requests", version: "0.1.0" });

const RequestSchema = z.object({
  type: z.enum(["refill", "appointment", "records", "question"]),
  patient_name: z.string(),
  contact: z.string(),
  details: z.string().optional()
});

// ... implementation details

C. Add scripts in package.json

"scripts": { "dev": "ts-node src/server.ts" }

D. Run it

npm run dev

Why this works: The server exposes two MCP tools that any MCP-aware assistant can call. The assistant stays the UI; the server is your guard-railed gateway to actions/data.

Edit the Claude config:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "patient-requests": {
      "command": "npx",
      "args": ["-y", "ts-node", "src/server.ts"],
      "cwd": "/absolute/path/to/mcp-patient-requests"
    }
  }
}

Restart Claude Desktop. You should see patient-requests available as a tool.

Try in Claude:

"Use patient-requests → create_request for a refill.
Name: Jane Doe. Contact: jane@example.com
Details: Needs 30-day refill of levothyroxine 75 mcg; last seen 3 months ago."

Then:

"Use patient-requests → list_requests (open). 
Summarize what needs staff action and draft a response for each."

Pattern: The assistant does the reasoning + drafting; the MCP tool logs/queues. Staff reviews and sends.

• FAQ search (no PHI): Add search_faq(query) that loads a local faq.json
• Scheduling preview (read-only): Add get_slots(provider, days_ahead) that reads shared calendar
• Refill safety checklist: Add refill_policy_check(med) to return clinic's checklist
• Status updates: Add update_request(id, status) to close out items

• No direct PHI to vendors unless you have BAAs; keep this server local or VPC
• Least privilege: expose read-only versions first (FAQ, slots)
• Logging: keep a local audit log of tool calls + inputs
• Config hygiene: Claude Desktop reads a simple JSON config—limit each server's access
• Stay current on MCP security guidance as adoption grows

• Single "front door" for patient requests that your AI can see and act on
• Faster triage: assistant drafts + MCP logs → staff approves
• Consistent policy adherence via checklists the assistant pulls in every time
• Searchable trail of requests (JSON/CSV) for QA and ops metrics