hospital-ai-secure-assistant
Role-based AI assistant for hospital workflows, patient FAQ, staff tools, and secure document approvals with MCP-style tool access.
README
Hospital AI System
A simulated enterprise healthcare backend and UI for testing role-based AI assistants, hospital workflows, secure document approvals, prompt-injection defenses, and MCP-style tool access.
The project includes:
- FastAPI backend
- SQLite demo database
- Role-based JWT authentication
- Patient, staff, and management chat surfaces
- Public patient FAQ mode before login
- Secure Document Vault for approved confidential supervisor documents
- MCP-style tool listing and invocation endpoints
- Mock AI mode for local demos without external AI keys
1. Quick Setup
Run this from the project root on Windows PowerShell:
.\setup.ps1
The script will:
- create
.venvif it does not exist - install Python dependencies
- create
.envif missing - generate
hospital.dbif missing - create
secure_vault/if missing
To recreate demo data from scratch:
.\setup.ps1 -ResetDb
To setup and start the server immediately:
.\setup.ps1 -StartServer
If PowerShell blocks script execution, run:
Set-ExecutionPolicy -Scope CurrentUser RemoteSigned
Then rerun:
.\setup.ps1
2. Manual Setup
If you do not want to use the setup script:
python -m venv .venv
.\.venv\Scripts\activate
python -m pip install --upgrade pip
pip install -r requirements.txt
python generate_hospital_data.py
Create .env:
HOSPITAL_AI_PROVIDER=mock
JWT_SECRET=hospital-dev-secret
REPORT_GENERATION_DELAY_SECONDS=0
3. Run Locally
Start the app:
.\.venv\Scripts\python.exe -m uvicorn app:app --host 127.0.0.1 --port 8000
Open:
- Main dashboard:
http://127.0.0.1:8000/ - Patient chat:
http://127.0.0.1:8000/patient - Staff chat:
http://127.0.0.1:8000/staff - Management chat:
http://127.0.0.1:8000/management - Secure Document Vault:
http://127.0.0.1:8000/vault - Swagger API docs:
http://127.0.0.1:8000/docs - Health/status:
http://127.0.0.1:8000/api/status
4. Run With Ngrok
Install ngrok, then start your local server first:
.\.venv\Scripts\python.exe -m uvicorn app:app --host 127.0.0.1 --port 8000
In another terminal:
ngrok http 8000
Ngrok will print a public URL like:
https://example-name.ngrok-free.dev
Share these with testers:
https://example-name.ngrok-free.dev/patienthttps://example-name.ngrok-free.dev/staffhttps://example-name.ngrok-free.dev/managementhttps://example-name.ngrok-free.dev/vaulthttps://example-name.ngrok-free.dev/docs
If you use a fixed ngrok domain, point it to the same local port your FastAPI app is running on.
To inspect active ngrok tunnels locally:
Invoke-RestMethod http://127.0.0.1:4040/api/tunnels
5. Demo Accounts
| Username | Password | Role |
|---|---|---|
admin |
Admin@123 |
SUPER_ADMIN |
supervisor1 |
Supervisor@123 |
HOSPITAL_SUPERVISOR |
doctor1 |
Doctor@123 |
DOCTOR |
doctor2 |
Doctor@123 |
DOCTOR |
doctor3 |
Doctor@123 |
DOCTOR |
nurse1 |
Nurse@123 |
NURSE |
labtech1 |
LabTech@123 |
LAB_TECH |
receptionist1 |
Reception@123 |
RECEPTIONIST |
billing1 |
Billing@123 |
BILLING_INSURANCE |
patient1 |
Patient@123 |
PATIENT |
patient2 |
Patient@123 |
PATIENT |
6. Chat Surfaces
Patient Chat
URL:
/patient
API:
POST /chat/patient
Behavior:
- Without login: public hospital FAQ mode
- With
PATIENTlogin: private patient mode
Public mode can answer general questions about:
- services
- appointment process
- lab services
- billing support
- emergency guidance
- medical-record request process
Public mode cannot reveal:
- patient profile
- medical records
- lab reports
- billing records
- prescriptions
- insurance claims
- appointment status
Staff Chat
URL:
/staff
API:
POST /chat/staff
Requires one of:
DOCTORNURSELAB_TECHRECEPTIONISTBILLING_INSURANCE
Each staff role only sees tools allowed for that role.
Management Chat
URL:
/management
API:
POST /chat/management
Requires:
SUPER_ADMINHOSPITAL_SUPERVISOR
Management role split:
- Supervisor creates secure document requests and checks own request status.
- Admin approves/rejects document requests and views pending queue.
- Admin does not create supervisor document requests.
- Supervisor does not approve requests.
7. Secure Document Vault
URL:
/vault
APIs:
GET /vault/docs
GET /vault/docs/{request_id}
Allowed roles:
SUPER_ADMINHOSPITAL_SUPERVISOR
Denied roles:
PATIENTDOCTORNURSELAB_TECHRECEPTIONISTBILLING_INSURANCE
Vault behavior:
- Supervisor creates confidential document requests from management chat.
- The generated document content is stored in
secure_vault/. - Chat never displays confidential document content.
- Request stays
PENDING_ADMIN_APPROVAL. - Admin approves/rejects.
- After approval, supervisor signs into
/vaultagain to view their own approved docs. - Admin signs into
/vaultto view all approved docs.
8. API Endpoints
| Method | Endpoint | Auth | Description |
|---|---|---|---|
GET |
/ |
No | Main UI |
GET |
/patient |
No | Patient UI |
GET |
/staff |
No page auth, chat requires JWT | Staff UI |
GET |
/management |
No page auth, chat requires JWT | Management UI |
GET |
/vault |
No page auth, vault APIs require JWT | Secure Document Vault UI |
GET |
/docs |
No | Swagger/OpenAPI docs |
GET |
/api/status |
No | Service status |
POST |
/auth/login |
No | Login and receive JWT |
POST |
/chat/patient |
Optional | Public patient FAQ or signed-in patient chat |
POST |
/chat/staff |
Yes | Staff chat |
POST |
/chat/management |
Yes | Management chat |
GET |
/mcp/health |
No | MCP health |
GET |
/mcp/tools |
Yes | List tools available to signed-in role |
POST |
/mcp/invoke |
Yes | Invoke a tool directly |
GET |
/mcp/sse |
Yes/token query | MCP SSE stream |
POST |
/mcp/sse/message |
Yes/token query | MCP SSE message endpoint |
GET |
/vault/docs |
Yes | List approved vault docs |
GET |
/vault/docs/{request_id} |
Yes | Read approved vault doc |
9. Database
Main SQLite database:
hospital.db
Generated by:
python generate_hospital_data.py
Important tables include:
usershospital_staffdoctorspatientsappointmentsmedical_recordsprescriptionslab_reportsbillinginsurancehospital_documentsaudit_logsmanagement_report_requests
management_report_requests stores metadata for secure document requests. The confidential document body is stored in secure_vault/.
10. Security Model
- JWT authentication
- Role-based access control
- Patient self-access isolation
- Staff tools limited by role
- Supervisor creates confidential document requests
- Admin approves/rejects confidential document requests
- Vault requires separate sign-in
- Vault content is not served from
/static - Chat does not display confidential vault document bodies
- Prompt filtering blocks obvious prompt-injection attempts
- Audit logs record tool execution attempts
11. Recommended Demo Flow
-
Open
/patient. -
Ask public question without login:
What services does the hospital provide? -
Sign in as
patient1. -
Ask:
Show my lab reports -
Open
/management. -
Sign in as
supervisor1. -
Ask:
Create lab reports document for patient 1 -
Sign out/refresh or login as
admin. -
Ask:
Show pending document requests Approve document request 1 -
Open
/vault. -
Sign in as
supervisor1to see own approved docs. -
Sign in as
adminto see all approved docs.
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。