A plain-English breakdown of where an LLM sits in the system, how each strategy decides to buy and sell, and how you can verify every claim on this page for yourself. No black boxes.
A large language model reads your natural-language prompt (“I want a cautious BTC bot”) and turns it into a JSON config that picks one of our predefined strategies and one of four personalities.
The same model powers agent-to-user chat, so you can ask your agent what it’s doing and get a reply in character.
What it does not do: decide trades. The LLM has no path to the exchange. It cannot place orders, move funds, or change the strategy code.
Five hand-coded strategies compute standard technical indicators (RSI, EMA, Bollinger Bands, ATR, volume ratios) and emit BUY/SELL signals when hardcoded threshold conditions are met.
Same market data in, same signal out. No hidden state, no model weights, no training loop. Every gate and threshold is visible in the strategy source files.
Every emitted signal ships with a reason string: exactly which conditions fired and what confidence score they produced.
Each strategy is a small, auditable Python file. Rules are listed in plain English here — full source in the repo.
Every signal a strategy produces passes through a separate risk engine before the trade reaches the exchange. The engine checks per-user exposure limits, monthly drawdown, cooldowns after consecutive losses, and account-level circuit breakers.
The behavior is fail-safe: if the risk engine is unreachable for any reason, the trade is blocked, not allowed through. The default is refusal to trade, never speculative execution.
After every executed trade, a post-trade check runs and can trigger cooldowns or risk-locks that will pause future entries.
Every strategy in worker/strategies/ is roughly 300 lines. All thresholds, gates, and confidence scoring are visible in the code. There is nothing else.
Every trade in your history includes the reason string that produced it: which conditions fired, current RSI, volume vs. average, confidence score. No decisions are hidden.
Paper mode is the default. It runs the full evaluate → risk-check → execute pipeline against simulated fills, so you can watch strategy behavior on your own account without a dollar at risk.