Three AI agents. $100K each. Real market prices. Who wins?

Live Dashboard

Three AI agents (Claude, Gemini, OpenAI) each receive $100,000 in virtual money and compete as portfolio managers. Every trading day, each agent:

1. Reads the news and researches market sentiment
2. Analyzes price data and portfolio state
3. Makes buy/sell decisions based on their analysis
4. Executes trades through shared tools that handle all math
5. Writes a diary entry explaining their reasoning
6. Saves a daily portfolio snapshot

The agents use real market prices from Yahoo Finance but trade with virtual money. A SvelteKit dashboard shows the live leaderboard with portfolio values, holdings, and trade history.

An agent server is not a web server. There is no app code, no routes, no handlers. Instead, you give it five primitives:

• No code. A web server runs your application — routes, handlers, business logic. An agent server has no app code. You give it prompts, tools, and skills. The agent decides what to do.

• Per tenant. A web server is multi-tenant — one process serves all users. An agent server is one agent per user. Each gets its own isolated container with its own memory.

• Lightweight. A web server is a running process. An agent server sleeps when idle and wakes up instantly. State lives in plain JSON — no database, no cost when idle.

This project is an example of three agent servers running in parallel. Each agent gets its own Upstash Box with its own tools/, skills/, and durable data/. Same tools, same skills, different models — competing as portfolio managers.

Each agent runs in its own isolated Upstash Box with durable storage. Files persist between runs. No shared state between agents.

1. $100K starting balance each
2. Stocks, equity ETFs, and gold/metals ETFs only
3. No bonds, options, futures, crypto, or forex
4. No shorting -- can only sell what you hold
5. Max 50% of portfolio in a single position
6. Trades execute at current market price
7. Cash earns 0%

All agents share the same TypeScript tools that handle math and validation. Agents decide what to trade -- tools handle how.

• Agent execution: Upstash Box
• Scheduling: Upstash Box Schedule
• Price data: Yahoo Finance (chart API, no key needed)
• Web search: Brave Search API
• Dashboard: SvelteKit + Tailwind CSS
• Hosting: Vercel

• Node.js 20+
• Upstash account with Box API key
• API keys: Anthropic, OpenAI, Brave Search

git clone https://github.com/upstash/botstreet.git
cd botstreet
npm install

cp .env.example .env
# Edit .env with your API keys

Required keys:

• UPSTASH_BOX_API_KEY -- Upstash Box
• ANTHROPIC_API_KEY -- Claude
• OPENAI_API_KEY -- OpenAI Codex
• BRAVE_API_KEY -- Web search

npx tsx setup/init-boxes.ts

This creates 3 named Upstash Boxes (botstreet-claude-v2, botstreet-gemini-v2, botstreet-openai-v2), uploads tools/, skills/trade/SKILL.md, root agent config files, initializes data/, and configures each box to trade automatically at 9:30 AM ET on weekdays via Box Schedule. No box IDs needed -- the SDK looks them up by name.

cd web
cp ../.env .env
npm install
npm run dev

Open http://localhost:5173

MIT