Someone Open-Sourced a Trading Agent That Touches 1,000 Markets
CloddsBot picked up 299 stars in a day to reach 1,589, and the description is either the most exciting or the most alarming thing on GitHub Trending depending on your risk tolerance. An autonomous trading agent built on Claude, running across Polymarket, Kalshi, Betfair, Binance, Bybit, Hyperliquid, Solana DEXs, and five EVM chains. MIT licensed. Built in twelve days during the Colosseum Agent Hackathon on Solana.
The architecture is the interesting part, not the trading. 119 bundled skills, 118 strategies behind a unified strategy engine, and connections to 21 messaging platforms including Telegram, Discord, and a built-in WebChat. You talk to it, it trades. There is real risk plumbing in there too: VaR and CVaR, circuit breakers, backtesting, whale tracking, arbitrage detection, copy trading. 446 commits, 253 forks, 21 open issues.
This is what an agent looks like when the domain has an unforgiving scoreboard. Nobody has to argue about whether the coding agent produced good code; here the loop either makes money or it does not, and every skill is one API call away from a real position. That is also why the skills count matters more than the strategy count. A hundred and nineteen skills is a hundred and nineteen ways to be wrong with real money, and prediction markets plus perps plus DEX routing is about as much blast radius as a hobby project can accumulate.
Code at https://github.com/alsk1992/CloddsBot
Worth saying plainly: a twelve-day hackathon project with a thousand markets wired up is a research artifact, not a product you point at your savings. What makes it worth reading is that it is one of the few fully open agent systems where the reward signal is unambiguous, which makes it a much better place to study skill design than anything with a human grading the output.
← Back to all articles
The architecture is the interesting part, not the trading. 119 bundled skills, 118 strategies behind a unified strategy engine, and connections to 21 messaging platforms including Telegram, Discord, and a built-in WebChat. You talk to it, it trades. There is real risk plumbing in there too: VaR and CVaR, circuit breakers, backtesting, whale tracking, arbitrage detection, copy trading. 446 commits, 253 forks, 21 open issues.
This is what an agent looks like when the domain has an unforgiving scoreboard. Nobody has to argue about whether the coding agent produced good code; here the loop either makes money or it does not, and every skill is one API call away from a real position. That is also why the skills count matters more than the strategy count. A hundred and nineteen skills is a hundred and nineteen ways to be wrong with real money, and prediction markets plus perps plus DEX routing is about as much blast radius as a hobby project can accumulate.
Code at https://github.com/alsk1992/CloddsBot
Worth saying plainly: a twelve-day hackathon project with a thousand markets wired up is a research artifact, not a product you point at your savings. What makes it worth reading is that it is one of the few fully open agent systems where the reward signal is unambiguous, which makes it a much better place to study skill design than anything with a human grading the output.
Comments