Beeper: private, uncensored AI credit marketplace
Next.js 15 + Tailwind 4. Buy discounted LLM inference credits (filled from a liquidity order book, 5% fee) or sell unused credits.
npm install
cp .env.example .env.local # set NEXT_PUBLIC_PRIVY_APP_ID
npm run dev # http://localhost:3000
npm run build && npm start
Sign in (Privy)
Login uses Privy with Google, email and wallet. Create an app at dashboard.privy.io,
enable Google / Email / External wallets under Login methods, add your domain (and http://localhost:3000)
under Allowed origins, then set NEXT_PUBLIC_PRIVY_APP_ID. Without it the site runs with sign-in disabled.
POST /api/orders requires a Privy access token, verified server-side against Privy's JWKS (lib/auth.ts).
Database
Postgres stores users, balances (integer micro-dollars), an append-only ledger, API keys (SHA-256 hashes only),
orders, usage logs and listings. Schema lives in db/migrations/*.sql; npm run build applies pending
migrations automatically when DATABASE_URL is set (or run npm run db:migrate).
On Vercel: Storage → Create Database → Neon (Postgres) and connect it to the project: that sets
DATABASE_URL. Supabase works too (use the pooled connection string). Then redeploy.
Payments:
- USDG on Robinhood Chain (chain 4663): each order asks for a unique exact amount (total + sub-cent suffix)
sent to
NEXT_PUBLIC_PAYMENT_ADDRESS. The buyer pays from their Privy-connected wallet or any wallet/exchange;
POST /api/orders/:id/confirm verifies the Transfer on-chain (by tx hash or by scanning logs) and credits once.
- Card via Whop:
POST /api/orders creates a Whop checkout; the signed payment.succeeded webhook
(/api/payments/whop) re-fetches the payment from Whop, checks order/amount and credits once.
PAYMENTS_MODE=test settles orders instantly: local testing only.
Local dev without Privy:
DATABASE_URL=postgres://... DEV_AUTH_BYPASS=1 NEXT_PUBLIC_DEV_USER=did:privy:dev PAYMENTS_MODE=test npm run dev
AI gateway
POST /v1/chat/completions and GET /v1/models are OpenAI-compatible. Requests authenticate with a Beeper key
(sk-beeper-…), are forwarded to OpenRouter (OPENROUTER_API_KEY) with data_collection: deny, and are billed
per token at the catalogue price in lib/data.ts. max_tokens is capped to what the balance can pay for;
streaming responses are billed from the final usage chunk (or an estimate if the client disconnects).
Images
POST /v1/images/generations and the chat's image mode use the Venice API (VENICE_API_KEY), billed per image
(Venice price + margin). Prompts referencing minors are rejected for every model.
Structure
app/page.tsx: landing: hero chat demo, buy card, liquidity book, live buys, privacy tiers, model catalogue, API snippets, FAQapp/sell: list unused credits · app/dashboard: balance/usage (demo data)app/api/me · keys · orders · listings · quote: account & marketplace API (Postgres)lib/data.ts: models, liquidity book, fee & order-fill logiclib/db.ts · lib/billing.ts · lib/keys.ts · lib/auth.ts: database, ledger, API keys, Privy auth
Before production
- Build the actual OpenAI-compatible gateway (
/v1/chat/completions) that meters usage against keys and routes upstream.
- Write real Terms / Acceptable Use / Privacy pages, and check credit resale is allowed by each upstream provider's ToS.
