OpenRouter

A self-hosted LLM routing API that proxies requests across multiple AI providers — OpenAI, Anthropic, Google Gemini — through a single OpenAI-compatible endpoint. Built with Bun, Elysia, Prisma, and React.

What it does

• Routes chat completion requests to the cheapest or most available provider for a given model
• Tracks per-key credit usage and deducts based on token cost per provider
• Lets users manage API keys, top up credits, and browse available models through a dashboard

Stack

Layer	Technology
Runtime	Bun
API framework	Elysia
Database ORM	Prisma (PostgreSQL)
Frontend	React 19, Tailwind CSS v4, shadcn/ui
Monorepo	Turborepo

Apps

• apps/backend — auth, API key management, model catalog, payments (port 3000)
• apps/api — the actual LLM proxy endpoint (port 4000)
• apps/frontend — dashboard UI (port 3001)
• packages/db — shared Prisma client

Prerequisites

• Bun >= 1.2
• PostgreSQL database
• API keys for the providers you want to support (OpenAI, Anthropic, Google)

Getting started

git clone https://github.com/your-username/openrouter
cd openrouter
bun install

Set up environment variables. Create .env files in apps/backend, apps/api, and packages/db:

# packages/db/.env
DATABASE_URL=postgresql://user:password@localhost:5432/openrouter

# apps/backend/.env
JWT_SECRET=your-jwt-secret
DATABASE_URL=postgresql://user:password@localhost:5432/openrouter

# apps/api/.env
DATABASE_URL=postgresql://user:password@localhost:5432/openrouter
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
GOOGLE_API_KEY=...

Run database migrations:

cd packages/db
bunx prisma migrate deploy

Start all services:

bun dev

Or run a specific app:

turbo dev --filter=backend
turbo dev --filter=frontend
turbo dev --filter=api

API usage

The proxy is OpenAI-compatible. Point your existing client at http://localhost:4000:

curl http://localhost:4000/api/v1/chat/completions \
  -H "Authorization: Bearer sk-or-v1-yourkey" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-3-5-sonnet",
    "messages": [{ "role": "user", "content": "Hello" }]
  }'

The model field follows the format company/model-name. The backend resolves which provider to use and handles billing automatically.

Database schema overview

• User — email/password auth, credit balance
• ApiKey — per-user keys with credit tracking and enable/disable toggle
• Model + Company — model catalog
• Provider + ModelProviderMapping — maps models to providers with per-token pricing
• OnrampTransaction — credit top-up history
• Conversation — request logs

Project structure

.
├── apps/
│   ├── api/          # LLM proxy (Elysia, port 4000)
│   ├── backend/      # REST API (Elysia, port 3000)
│   └── frontend/     # Dashboard (React, port 3001)
└── packages/
    ├── db/           # Prisma schema + client
    └── ui/           # Shared React components

Development notes

• The frontend uses Eden Treaty for end-to-end type-safe API calls — no manual type definitions needed for client-server communication
• Provider selection for a model is currently random across all mapped providers. You can extend apps/api/src/index.ts to implement cost-based or latency-based routing
• Credits are stored as integers. The current formula is (inputTokens * inputCost + outputTokens * outputCost) / 10

License

MIT