Claude agent traded on Polymarket

A language model can read news, odds and blockchain records, turn them into a probability estimate, and place a trade. An ex-Anthropic engineer says he built that system for Polymarket with Claude and released the code publicly this week. (github.com) (x.com) Prediction markets let traders buy “yes” or “no” shares on future events, with prices moving like implied odds. Polymarket describes itself as a prediction market where users trade on politics, sports, crypto and other events. (polymarket.com) The project described by xMayeth uses retrieval-augmented generation, a method that feeds an artificial intelligence model outside documents before it answers. Polymarket’s own open-source agents framework lists support for retrieval-augmented generation, news providers, web search and market data connectors. (github.com) (x.com) In practice, that means the model is not guessing from memory alone. It can pull current articles, market prices and on-chain transaction data, then turn that evidence into a trade decision and send orders through Polymarket’s application programming interfaces. (github.com) (polymarket.com) The author said the agent hit 74 percent of its calls and made about $41,000 over 90 days. Those figures come from the author’s own report on X, and I could not independently verify wallet-level performance from the materials publicly surfaced in search results. (x.com) That caveat matters because Polymarket bots are already a crowded field, with public repositories advertising strategies from arbitrage to market making to short-term crypto bets. Several recent GitHub projects also claim high win rates, but those numbers are usually self-reported and often cover short test windows. (github.com 1) (github.com 2) (github.com 3) The new wrinkle is not that bots exist, but that a Claude-based setup is being packaged as a reusable framework for other traders. Polymarket itself has maintained an open-source agents repository since 2024, and its GitHub organization now also hosts tools and “agent skills” aimed at developers building automated traders. (github.com 1) (github.com 2) (github.com 3) Anthropic’s Claude is a general-purpose model, not a trading product, but its coding and tool-use features have made it popular for agent-style software that can search, read files and run scripts. Anthropic’s product pages say Claude can analyze data, write code and use web search, which fits the kind of workflow these Polymarket agents rely on. (claude.com) (anthropic.com) The result is a simple pitch with a hard-to-check edge: give a model fresh information, let it price an event faster than a human, and automate the order. Whether this specific bot keeps beating the market will depend less on the thread than on whether its public code still finds mispriced contracts after everyone else reads it. (x.com) (github.com)