the core idea
Programmable.
Composable.
AI-native.
We don't ask the LLM to guess buy/sell every tick.
We ask it to write the strategy, compile it with esbuild, and hot-reload at runtime.
.ts
strategies as code
git
version · branch · PR
Write once.
Run on any exchange.
Strategy code has zero knowledge of which exchange it runs on.
a strategy in ~10 lines
class MomentumStrategy extends BaseStrategy { readonly strategyId = 'momentum' async evaluate(context: StrategyContext) { const { symbol, threshold } = this.params.base const tick = context.getData('price', symbol) if (!tick || tick.price < threshold) return [] return [this.instruction('perp', 'placeOrder', { symbol, side: 'buy', type: 'market', amount: 0.01 })] } }
context.getData() — no exchange SDK. No WebSocket boilerplate.
this.instruction() — no direct API calls. Swap the exchange in one line.
This entire class can be generated by AI, compiled with esbuild, and hot-loaded at runtime.
architecture
Four layers.
Every one replaceable.
Swap any layer without rewriting the others. The AI layer is just one node in the graph.
Monitor ─▶ data collection (WS / poll → JSONL) TriggerManager ─▶ Cron + Monitor AND, time-windowed Strategy ─▶ AI inference / rule-based decision Executor ─▶ trade execution
monitor
WS + Poll
→ JSONL stream
trigger
AND-logic
time-windowed
strategy
LLM ✕ code
hot-reload
executor
multi-venue
paired orders
vs general agents
Trading-native by design.
🐋
OpenWhale
- · role — programmer
- · output — TypeScript code
- · audit — git diff
- · LLM call — on recompile
- · scope — vertical (trade)
🤖
OpenClaw · Hermes
- · role — agent runtime / loop
- · output — tool calls
- · audit — logs
- · LLM call — every tick
- · scope — horizontal
tl;dr — they ask AI to make a trade. We ask AI to write a trader.
in the wild
Real strategies, shipped.
copy-trading
Mirror any wallet, proportionally.
Monitor a target address. Mirror trades at a configurable ratio with position caps. One strategy, any exchange.
cross-exchange-arb
Hyperliquid ↔ Binance
Spread AND funding rate — both must align within 60s. Paired orders, two account instances.
llm-news-react
News → structured output → trade.
Schema-validated. Acts only on { confidence > 0.7, action != hold }.
on-chain events
Contract event → action.
Listen to smart contract events. Trigger on-chain or off-chain actions automatically.
nft sniping
Mempool → milliseconds.
Monitor launchpad events. Execute within milliseconds of a trigger.
join
Sign the manifesto.
5 steps · ~2 minutes · NFT raffle on completion.
