Work · 02 / 06 · 2025
open-verse
LLM trading agents compete on paper markets. An orchestrator ticks the market, calls the models, places paper orders, and a dashboard shows who is winning.
- Role
- Solo build
- Year
- 2025
- Stack
- TypeScriptNode.jsOpenAI SDKAlpaca Paper TradingBinance Spot Testnetbetter-sqlite3Postgres (optional)ZodPino
Problem
Claims about LLMs "trading" are usually a screenshot and a prompt. I wanted a controlled arena where several models make real decisions on real prices, with hard risk caps, and where the scoreboard is the only argument.
What I built
open-verse runs one or more named bots, each with its own LLM provider, model, and API key, and a cash balance. On a timer the orchestrator pulls prices, asks each model for a decision, and submits paper orders under position and per-order caps. State lives in SQLite with an optional Postgres mirror; a leaderboard API and a live dashboard show standings. It supports Alpaca paper trading by default and Binance Spot Testnet as an alternate venue. Nothing touches live money.
Safety rails (MAX_POSITION_USD, MAX_ORDER_USD, starting cash, decision interval, symbol universe) are configuration, so an experiment is an env file, not a code change.
Outcome
A reusable harness for comparing model behaviour on markets, with a leaderboard that is hard to argue with. Started in October 2025 and still evolving.