Support Arbiter - Multi Agent Customer Support

Next.js app with a Hono API and a simple multi agent AI system.

What it does

1. Stores users, conversations, messages, orders, deliveries, invoices, payments, refunds
2. Routes each incoming message to a specialized agent using an LLM router
3. Enforces tools first behavior in sub agents by calling DB backed tools before generating text

Tech

1. Next.js App Router UI
2. Hono API mounted at /api/*
3. Prisma with PostgreSQL
4. Vercel AI SDK with Google Gemini via @ai-sdk/google

Quick start

1. Install

npm install

1. Create .env from .env.example and set values

Required

DATABASE_URL="postgresql://..."
GOOGLE_GENERATIVE_AI_API_KEY="..."

Optional

GOOGLE_GENERATIVE_AI_MODEL="gemini-2.5-flash"

1. Database

npm run prisma:generate
npm run prisma:migrate
npm run prisma:seed

1. Run

npm run dev

API

1. GET /api/health
2. POST /api/chat/messages
3. GET /api/chat/conversations?userId=demo-user
4. GET /api/chat/conversations/:id
5. DELETE /api/chat/conversations/:id
6. GET /api/agents
7. GET /api/agents/:type/capabilities

Seeded examples

The seed creates a demo user and sample data.

Demo user

1. Email demo@acme.com
2. UI uses userId=demo-user and the server maps it to the seeded demo user id

Orders

1. Order number A10001 status SHIPPED carrier UPS tracking 1Z999AA10123456784
2. Order number A10002 status PROCESSING carrier USPS tracking 94001118992238569210

Invoices

1. Invoice number INV-10001 status PAID total 129.99 USD
2. Invoice number INV-10002 status OPEN total 49.99 USD

Refunds

1. Payment for INV-10001 has a refund row with status NONE

Test queries

Use these in the UI after seeding.

1. Where is my order A10001?
2. Check delivery status for A10002
3. Can I get invoice INV-10001?
4. Do I have a refund for INV-10001?
5. How do I reset my password?

If you see quota errors from Gemini, verify GOOGLE_GENERATIVE_AI_API_KEY or change your key.

Notes

• If you see an 'An error was encountered' or similar on the hosted app, please let me know chances are the api key is facing quota issues.
• The hosted version has rate limiting of 10 req/min so try to stay under that limit. If you need more for testing, then I can tweak the limit, just let me know.