PriorAuth AI Agent

Automates US healthcare prior authorization using a 3-tool MCP pipeline. A clinician types one sentence — the agent returns a complete PA submission in under 30 seconds.

Stack

• MCP servers: FastMCP (Streamable HTTP, MCP 2025-03-26 spec)
• LLM: Google Gemini API (gemini-2.5-flash, free tier)
• FHIR: SMART Health IT R4 sandbox with mock fallback
• Platform: Prompt Opinion A2A orchestrator
• Frontend: Next.js + Tailwind CSS dashboard

Architecture

Layer	Component	Port
MCP Tool 1	fhir-patient-fetcher	8001
MCP Tool 2	coverage-checker	8002
MCP Tool 3	justification-writer	8003
Orchestrator	FastAPI chain (Tool 1 → 2 → 3)	8000
Platform	Prompt Opinion A2A Orchestrator	—
Frontend	Next.js dashboard	3000

Quick Start

cd backend/prior-auth-agent
cp .env.example .env
# Add your GEMINI_API_KEY to .env
docker compose up --build

Verify the servers are running

curl http://localhost:8000/health   # orchestrator
curl http://localhost:8001/health   # fhir-fetcher
curl http://localhost:8002/health   # coverage-checker
curl http://localhost:8003/health   # justification-writer

Demo patients

ID	Patient	Scenario	Denial risk
mock-123	Sarah Johnson	RA — Humira (adalimumab)	Low
mock-456	Marcus Williams	Psoriasis — Skyrizi (risankizumab)	Medium
mock-789	Linda Okafor	T2D — Ozempic (semaglutide)	High

SHARP Context Headers

When called via Prompt Opinion, the orchestrator injects:

• X-Sharp-Patient-Id — patient identifier
• X-Sharp-Fhir-Token — FHIR OAuth2 bearer token
• X-Sharp-Fhir-Base-Url — FHIR server base URL

Project structure

backend/prior-auth-agent/
  services/
    fhir_fetcher/        MCP Tool 1 (port 8001)
    coverage_checker/    MCP Tool 2 (port 8002)
    justification_writer/ MCP Tool 3 (port 8003)
  shared/
    models.py            Pydantic v2 data contracts
    sharp_context.py     SHARP header utilities
  docker-compose.yml
  .env.example
frontend/                Next.js dashboard

Contributors

Name	Role
Vipin Bhatt	Backend MCP servers, FHIR client, coverage checker
Yashas Yadav	Frontend, orchestrator, platform integration