COMMENT "CODEBOT"Posted Oct 21, 2026
The 50-market arbitrage bot a 19-year-old built with Claude Code
The viral $68 to $750,000 bot is a cross-market arbitrage engine: Claude Code handles the mispricing logic, a Binance websocket feeds live BTC every second, and an execution layer fires the moment a gap opens across markets. Here is the architecture and the prompts to clone it.
01Quick steps
- 1Architecture: Claude Code = mispricing logic, Binance websocket = real-time BTC, execution layer = fires on a cross-market gap.
- 2Prompt Claude Code to build the scanner (prompt below), streaming prices across N markets and flagging gaps over a threshold.
- 3Wire a Binance websocket (wss stream) so BTC price updates feed the signal every second.
- 4Put the risk in code: max position size, a hard stop on every trade, auto-shutdown on low liquidity.
- 5Backtest with real fees and slippage, then run it 24/7 and monitor alerts only.
02Your ready to use prompt
Build an arbitrage scanner: stream live prices across N markets, flag any pair whose spread exceeds X percent after fees, and return the trade with size, entry, stop and target. Use a Binance websocket for real-time BTC. Log every signal.
03Full tutorial
You are cloning a three-part system: (1) Claude Code writes and holds the mispricing logic, (2) a Binance websocket streams real-time BTC into the signal, (3) an execution layer places and manages the trade when a cross-market gap opens. The edge is latency: it acts on spreads that close in seconds, so the whole thing is built for speed, not chart-reading.
Prompt: 'Build an arbitrage scanner in Python: stream live prices across N markets via websockets, compute the spread between correlated pairs, and flag any spread over X percent AFTER fees and slippage. Return {market, spread, size, entry, stop, target}.' Have Claude Code write it as separate modules (feed, signal, risk, execution) so you can test each in isolation.
Add a Binance websocket (wss://stream.binance.com) subscribing to the BTC trade/book stream, so your reference price updates every second instead of on a slow REST poll. Feed that price into the signal module. Latency here is the edge: a REST poll is too slow for gaps that close in seconds.
Set X (the minimum spread) ABOVE your round-trip cost: taker fees on both legs + expected slippage + a buffer. Below that you are trading noise. Have the scanner only emit a signal when the net edge after costs is positive. This single number is what separates profit from churn.
In the risk module: a fixed max position size, a hard stop on every trade, and an auto-shutdown that halts new entries when liquidity or depth drops below a floor (so you never fire into a market you cannot exit). These are code rules, not discretion, so the bot cannot override them at 3am.
On a cleared signal, place both legs (or the single directional trade) programmatically with the stop attached, and confirm fills. Handle partial fills and timeouts. If one leg fills and the other does not within your window, flatten immediately. Never leave a half-open arb.
Replay historical data through the same modules with actual fees and slippage applied. If the edge survives costs over a few weeks of data, it is real; if it only works at zero fees, it does not exist. Tune X and position size here, not live.
Deploy on a small always-on box near the exchange, start with tiny size, and watch the logs and alerts. Scale size only after it is consistently net-positive live. The creator said they only monitored notifications once it was running, that is the goal state.
The $750,000 is a self-reported, promotional number and the original bot runs on prediction markets. This guide is the generic arbitrage architecture, which is real and buildable. Only ever risk money you can afford to lose. This is not financial advice.
The $750,000 total is the creator's self-reported figure and the original runs on prediction markets; this is the generic architecture. Only trade money you can afford to lose. Not financial advice.