convez.ai
MCP ServerDesenvolvimento WebAvançado
niket-sharma

CareTriage: Tutorial de Claude Code com App de Triagem Médica

niket-sharma/claude-code-healthcare-tutorial

Tutorial prático: construa um sistema de triagem médica seguro com Claude Code

Ver no GitHub
InicialPluginScore · de 0 a 100

Um tutorial passo a passo para dominar Claude Code construindo uma aplicação real de triagem médica (CareTriage). Cobre permissões, hooks, MCP, subagentes, times de agentes, Ralph loops e padrões de segurança críticos para sistemas sensíveis. Inclui backend Python com FastAPI, detecção de red-flags determinística e frontend React.

Instalação

$npx convez-ai niket-sharma/claude-code-healthcare-tutorial

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
1
Contribuidores
1

Casos de uso

  • Aprender arquitetura de MCP servers através de um caso real
  • Entender padrões de segurança para sistemas de saúde assistidos por IA
  • Dominar agent teams, hooks e subagentes do ecossistema Claude
  • Implementar fail-safes determinísticos antes de chamadas LLM

Veredito

Um excelente exemplo de produção real que demonstra segurança em sistemas sensíveis: red-flags determinísticas, disclaimers obrigatórios, fallbacks inteligentes. Invaluável para quem constrói MCP servers críticos, mas é um tutorial educacional — nunca para uso clínico real.

README

A tradução para o português deste README ainda está na fila.

CareTriage

NOT A MEDICAL DEVICE. CareTriage is educational software built to demonstrate AI-assisted development workflows. It is not approved for clinical use, not a substitute for professional medical advice, and must never be deployed to real patients. Every response carries an explicit disclaimer to that effect.


What is CareTriage?

CareTriage is a symptom-intake and triage-suggestion service. A patient describes their symptoms; the service records a structured intake, runs a deterministic red-flag screen, and (if no emergency is detected) calls an LLM to produce a triage suggestion in one of four levels:

LevelMeaning
self_careMinor, self-limiting — manage at home
see_clinicianWarrants professional evaluation, not urgent
urgentSame-day or next-day care needed
seek_immediate_careCall 911 or go to an emergency room now

Urgent and immediate-care encounters are automatically added to a clinician review queue.


Safety Design

1 — Red-flags run first, always

Before any LLM call, a deterministic keyword screen (app/core/red_flags.py) checks for emergency indicators: chest pain, stroke signs, difficulty breathing, severe bleeding, anaphylaxis, suicidal ideation, sudden severe headache, and loss of consciousness. If any phrase matches, the response is seek_immediate_care and the LLM is never contacted.

2 — Fail-safe fallbacks

If the LLM call fails or returns malformed output, the service falls back to see_clinician with an instruction to consult a clinician — never to a silent failure or an empty response.

3 — Disclaimer on every response

Every triage result — in the API JSON and in the UI — carries:

"This output is not a medical diagnosis. Always consult a qualified clinician for medical advice."

4 — Synthetic data only

No route, test fixture, seed script, or log statement may accept, store, or emit real patient health information. All sample data uses obviously fictional names, dates, and symptoms.

5 — Secrets stay in the environment

The OpenAI API key is read from .env at runtime via python-dotenv. It never appears in source files, test files, or CI config.


Features

  • POST /triage — submit age, sex, and symptoms; receive a triage level, rationale, and safety disclaimer
  • GET /encounters — paginated list of all recorded encounters for the clinician dashboard
  • Automatic review queue — urgent and seek_immediate_care encounters are enqueued atomically with the encounter write
  • Red-flag module — 8 emergency categories, 60+ trigger phrases, zero network dependency
  • React dashboard — clinician-facing encounter list (Vite + TypeScript)
  • Full test suite — 92 tests covering triage logic, red-flag detection, review queue behaviour, and API endpoints

Setup

Prerequisites

  • Python 3.11+
  • Node.js 18+
  • An OpenAI API key

1 — Clone and configure

git clone <repo-url>
cd claude-code-healthcare-tutorial

cp .env.example .env
# Edit .env and set:  OPENAI_API_KEY=<your-key>

2 — Backend

python3 -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

pip install -r requirements.txt

uvicorn app.main:app --reload --port 8000

The API is now available at http://localhost:8000.
Interactive docs: http://localhost:8000/docs

3 — Frontend

cd frontend
npm install
npm run dev

The dashboard is now available at http://localhost:5173.
API calls are proxied to localhost:8000 automatically.


Development

# Run all backend tests
pytest

# Run tests with coverage
pytest --cov=app --cov-report=term-missing

# Lint and format (Python)
ruff check .
ruff format .

# Frontend tests and lint
cd frontend
npm test
npm run lint

Project Structure

app/
  api/          # FastAPI routers
  core/         # config, DB engine, red-flag module
  models/       # SQLAlchemy / SQLModel ORM models
  services/     # business logic, OpenAI client
tests/          # pytest suite
frontend/       # React + Vite dashboard
docs/           # architecture notes

License

Educational use only. See CLAUDE.md for the full project constitution and safety rules.

Ver o original no GitHub

Bom para

Relacionados