Claude bot: Citadel‑style MM
A Claude‑built bot demomed a Citadel‑style market‑making strategy that captured spreads on BTC/ETH pairs and reported a $234 profit in the video — it emphasized spread capture over directional bets and referenced VPIN kill switches and mean‑reversion logic. The clip is being touted as a concrete project idea for replicable market‑making risk models (x.com).
A recent hands‑on case study documented a Claude Code workflow that produced a 27‑file, production‑grade crypto trading system over 14 sessions and 961 tool calls, illustrating the scale of orchestration used in community demos. (explore.n1n.ai) Production-ready market‑making repositories explicitly include multi‑level kill switches and VPIN/toxicity detectors — the mg‑market‑maker project lists a “5‑level kill switch” and VPIN toxicity detection in its feature set. (github.com) Mainstream market‑making firms have signaled institutional entry into crypto: Citadel Securities was reported by CoinDesk and Bloomberg as exploring crypto market‑making and liquidity provision on major exchanges. (coindesk.com) Several public examples show Claude and similar agents being used end‑to‑end: a Polymind/Claude trading‑bot repo on GitHub and a Moon Dev YouTube walkthrough both publish code, backtests, and deployment notes for AI‑constructed bots. (github.com) For data and plumbing, the CCXT library is the de‑facto Python/JS interface to >100 exchanges for live execution and backtests, while Binance publishes daily/monthly public trade files and REST/WebSocket endpoints for historical and real‑time trades. (github.com) Operational metrics to reproduce and stress‑test the demo should include VPIN/toxicity thresholds (industry guides flag VPIN >75% as high toxicity), spread‑capture rate, inventory variance, execution slippage, and PnL attribution — tools like VisualHFT surface VPIN and order‑book imbalance and mg‑market‑maker includes PnL attribution modules. (thrive.fi) If replicating in code, start from a canonical Avellaneda‑Stoikov implementation (papers, GitHub notebooks, and Hummingbot’s Avellaneda strategy exist) and add a VPIN stream plus configurable multi‑level kill switches for toxicity and latency breaches. (math.nyu.edu)
