Production-grade algorithmic trading infrastructure for Node.js — from Pine Script execution and LLM signal generation to live exchange connectivity and interactive dashboards. One codebase for backtest, paper, and live trading. Telegram notifications included. UI Dashboard included.
Start here: Clone the reference implementation — a fully working news-sentiment AI trading system with LLM forecasting, multi-timeframe data, and a documented February 2026 backtest.
# Scaffold a new AI agent trading project
npx -y @backtest-kit/cli --init --output my-trading-bot
cd my-trading-bot
npm start| Package | Description |
|---|---|
backtest-kit |
Core engine — temporal context via AsyncLocalStorage, exchange adapters, risk management, signal lifecycle |
@backtest-kit/ui |
Full-stack dashboard — candlestick charts, signal tracking, PnL analytics, trailing stops visualization |
@backtest-kit/signals |
50+ technical indicators across 4 timeframes, order book depth, AI-ready markdown reports |
@backtest-kit/ollama |
Multi-provider LLM wrapper (OpenAI, Claude, DeepSeek, Grok, Mistral, Ollama, 10+) with structured output |
@backtest-kit/pinets |
Run TradingView Pine Script v5/v6 locally — 1:1 syntax, 60+ built-in indicators |
@backtest-kit/graph |
Typed DAG execution for multi-timeframe strategies — parallel source nodes, serializable to DB |
@backtest-kit/cli |
CLI for scaffolding AI agent trading projects with CLAUDE.md skill contracts |
Look-Ahead Bias is Architecturally Impossible
Temporal context flows automatically through AsyncLocalStorage — every getCandles() call is implicitly bounded to the current backtest tick. No timestamp parameters to pass, no future data to accidentally leak.
Same Code, Backtest → Paper → Live
Strategy functions are unaware of execution mode. Switch from historical simulation to real exchange connectivity by changing one flag.
LLM as Signal Generator, Not as Indicator Replacement
The framework provides structured pipelines for injecting multi-timeframe technical analysis, order book data, and news sentiment into LLM context — while keeping temporal isolation and validation invariants intact.
A series of in-depth articles documenting the engineering decisions behind the framework:
- Look-Ahead Bias — How
AsyncLocalStoragemakes temporal contamination architecturally impossible - Second-Order Chaos — Why thousands of identical bots trade against themselves at a loss
- Claude Trader — How AI gets hands for autonomous strategy iteration via Claude Code
- Option Hedging — Why the price drops in a single candle and how to adapt
- AI Strategy Workflow — AI workflow for identifying and updating liquidation cascade criteria
- AI Strategy Blueprint — ReAct-pattern LLM trading agent with live test results
- AI News Trading Signals — News sentiment analysis as a trading signal: methodology and backtest
MIT © tripolskypetr