Sardine: A stock market where AI agents trade in real time

Sardine is an AI-driven market simulation platform built with Mastra, TanStack Start, and a live trading-style dashboard. Sardine combines a configurable market simulator, Mastra-powered trading and research agents, a what-if chatbot, and a multi-panel dashboard for watching price action, order flow, and agent behavior in real time. Sardine is designed for running market simulation sessions that feel closer to a trading terminal than a toy demo: - Launch a session with configurable symbol count, agent count, tick timing, and trader mix. - Run Mastra trading, research, and chatbot agents against a live simulation loop and typed tools. - Stream watchlists, candlesticks, order book depth, time and sales, research notes, and agent events to the dashboard. - Bootstrap from Alpaca market data when credentials are available, or fall back to local seed prices for offline development. - Explore a curated local dev profile today, while keeping full S&P 500 ticker metadata scaffolded in code for larger-scale experiments. - Session-based simulations: each run persists its own symbols, tick settings, agent roster, and runtime state. - Mastra agent stack: shared trading, research, and chatbot agents adapt behavior through request context and tool calls. - Matching-engine workflow: orders flow through a limit-order-book engine, then into portfolio reconciliation and persistence. - What-if interaction loop: the chat API can inject events into a live session and observe how the market reacts over time. - Realtime monitoring: a separate WebSocket server broadcasts market, runtime, and agent activity to the dashboard. flowchart LR subgraph UI["Frontend (TanStack Start)"] Landing["Routes: / and /dashboard"] Session["Session dashboard: /dashboard/$sessionId"] ChatUI["Chat panel and dashboard widgets"] end subgraph App["App server"] Loaders["Session loaders and route handlers"] ChatAPI["/api/chat"] end subgraph Runtime["Simulation runtime"] Runner["src/server/sim-runner.ts"] Orchestrator["SimOrchestrator + SimClock"] Registry["AgentRegistry + batch scheduler + autopilot"] Match["MatchingEngine + limit order books"] WS["WebSocket server (:3001)"] end subgraph Mastra["Mastra"] Trading["tradingAgent"] Research["researchAgent"] Chatbot["chatbotAgent"] Tools["Tools: market data, portfolio, order, event injection, sim query, firecrawl, wait-and-observe"] end subgraph Data["Persistence and external services"] PG["Postgres + Drizzle + Mastra storage"] Alpaca["Alpaca market bootstrap"] Firecrawl["Firecrawl research scraping"] end Landing --> Loaders Session --> Loaders ChatUI --> ChatAPI ChatAPI --> Chatbot Chatbot --> Tools Loaders --> PG Session |"ws://localhost:3001"| WS Runner --> Orchestrator Runner --> Registry Orchestrator --> Match Runner --> PG Runner --> WS Registry --> Trading Runner --> Research Trading --> Tools Research --> Tools Tools --> PG Tools --> Alpaca Tools --> Firecrawl - A session is created from the dashboard and stored in Postgres with its symbol set, timing, and trader distribution. bun run sim picks up pending sessions, bootstraps prices from Alpaca when available, and falls back to local seeded market state otherwise.- Bootstrap logic creates the initial books, bars, portfolios, and namespaced agent instances for the session. - On each tick, the orchestrator advances simulated time, runs the active LLM cohort, applies autopilot directives for the rest, and releases research into the market. - Orders are processed through the matching engine, portfolio state is reconciled, and the updated session state is persisted. - The WebSocket server broadcasts runtime state, price bars, order book snapshots, trades, and agent events back to the dashboard. - The chatbot can inject events and query session state through /api/chat , closing the loop between operator input and simulated market behavior. git clone https://github.com/lightbearco/sardine.git cd sardine bun install cp .env.example .env.local Start with the values in .env.example , then add the additional variables called out in Environment. Sardine uses Postgres via Drizzle and Mastra storage. A Neon database works well for local development. bun run db:push Run the frontend and simulation worker together: bun run dev:full Or run them in separate terminals: bun run dev bun run sim - App UI: http://localhost:3000 - Realtime WebSocket server: ws://localhost:3001 Open /dashboard , create a session, and launch the simulation from there. | Command | Description | |---|---| bun run dev | Start the TanStack Start app on :3000 | bun run dev:full | Start the app and simulation worker together | bun run sim | Start the simulation runner and WebSocket server | bun run build | Build the production app | bun run preview | Preview the production build | | Command | Description | |---|---| bun run test | Run the full Vitest suite | bun run test:trading:fast | Run fast trading-agent and tool tests | bun run test:trading:live | Run the live trading-agent smoke test with real mod