3 Free GitHub Repos That Print Money with AI Trading: Freqtrade, Vibe Trading and NOFX

Why does building a trading bot from scratch still feel intimidating when GitHub already hosts polished execution engines for free? Miles Deutscher flips the question: the barrier is not technical, it is choice. Over the past weeks he scanned the big hubs under finance, quantitative finance, stock market, crypto and trading-bot tags and distilled three repos that actually execute. The audience is clear: builders who want a plug-and-play start now but plan to refine a custom strategy later. The backbone is the trio — Freqtrade, Vibe Trading and NOFX — each filling a different gap. Think of it as LEGO: separate bricks, one coherent system when combined.

Repo one is Freqtrade. A free and open source crypto trading bot written in Python, matured since 2017 with a large Discord community and thorough docs. It supports both spot and futures via CCXT (the exchange-connectivity library that lets one codebase talk to many venues), including Binance, Bybit, OKX, Hyperliquid (a decentralized perpetual venue) and Kraken. Control comes through two doors: web UI and Telegram. Inside are backtesting, plotting, money management and machine-learning strategy optimization. The loop is simple: 1) Define a strategy, 2) Backtest on history, 3) Validate with a dry run (forward test on live market data without real money), 4) Go live if it survives. It is a closed workshop in a single file.

What makes Freqtrade timely is a context shift. The video notes that crypto venues like Binance and Hyperliquid now list stocks, commodities, FX and tokenized real-world assets (RWA — equities and commodities represented on-chain). That widens the mandate beyond crypto; you can trade stock-like exposures through the same API. Versus classic brokers, the entry friction is lower and the APIs are cleaner for AI agents. On Telegram you can use /performance to see results, /start to launch a strategy, /stop to halt a run and spin up several bots in parallel chats. The sharpest advice: a backtest may look dazzling, but the live dry run is the real exam. As a demo he asks Claude to backtest a Bitcoin and Ethereum breakout (a breakout — price punching through resistance) strategy, then watches entry and exit orders and the dashboard positions in the web UI. Setup was delegated to Claude Code. The catch: the bot lives in a Python terminal, so it must stay on 24/7, which means a VPS (virtual private server — a remote computer that never sleeps) is needed; Hostinger at a few dollars is suggested. For live trading, Claude's inference layer talks to the exchange through MCP (model-context protocol — the LLM commanding an external tool) to route orders.

Repo two is Vibe Trading. It does the opposite of Freqtrade: research, not execution. Tagline is turn market questions into runnable research, positioning itself as a personal trading agent. The self-improving agent takes a theory and forges it into a backtest-ready strategy. Under the hood are 69 skills, 29 swarm presets (a swarm — specialists for macro, quant, risk, catalyst writing a joint report), 21 tools and 5 data sources. Capabilities are crisp: ask a trading question, backtest an idea, review your trades, ingest institutional filings and fund letters, run analyst teams. Where Freqtrade runs a strategy, Vibe Trading invents one. For execution it can connect to Alpaca (a US broker offering paper trading for stocks and crypto) and other venues, but analysis is the core. Its Alpha Library holds 450+ prebuilt quant alphas (an alpha — a formula that explains price) across four zoos: Qlib158, Kakushadze 101, GTJA191 and an academic set. Like lab kits on a shelf: grab, test, refine.

The demo makes the research power tangible. He asks for the best momentum (momentum — the bet that winners keep winning) strategy across Nvidia, Meta, Apple, Microsoft and Amazon over the last two years, backtested with realistic fees and plotted against buy and hold, then deployed as paper trades on an Alpaca account with $100,000. A dashboard appears: total return, drawdown (peak-to-trough give-back), trades, positions and attribution. Orders land on the paper exchange and are visible on the board. The video adds an honest aside: critically, this is not a perfectly engineered strategy example, just a flow example. The real strength is scanning the Alpha Zoo: benchmarking 462 strategies against the S&P 500 with one command and flagging Alpha 10132 as a candidate that passes the out-of-sample test, albeit with a small edge versus buy and hold. Takeaway: the best use is to generate and prune, then watch months in paper.

Repo three is NOFX, the video's favorite. Described as an AI trading terminal assistant for US stocks, commodities, forex and crypto. NOFX is the execution middle layer between language model and exchange. Its standout is Autopilot: fund a wallet and strategies run on the server across Hyperliquid and centralized exchanges without keeping a local terminal open. Picture three layers: 1) Trading Terminal (React + TradingView) — dashboard, Strategy Studio, competition leaderboard, 2) API Server (Go) — JWT auth and encrypted credential store, 3) Trader Runtime — strategy engine, Autopilot and risk engine split. Every trader loops: read market structure, decide, execute, record reasoning. A Go runtime clamps every order with hard risk limits the model cannot override. Nine exchanges are supported: Binance, Bybit, OKX, Hyperliquid, Bitget, KuCoin, Gate, Aster and Lighter. On Hyperliquid the same runtime also trades tokenized US equities, commodities, indices, FX and pre-IPO perpetuals such as TSLA, NVDA, GOLD, SPX and EUR. Everything runs on your own machine; credentials are encrypted at rest and never leave it.

Visually, NOFX differs. The terminal shows a cost and liquidation heatmap (density of where leveraged positions would liquidate), order book, signal matrix and orchestration topology side by side. On the strategy side the Signal Board shows live directions; in the video Bitcoin, VVV, Ethereum and Hyperliquid tilt long with direction history and a heatmap below. Creating a strategy means opening Strategy Studio and setting style presets, coin universes, indicators, leverage, entry confidence and a custom prompt. You can also deploy prebuilt strategies with one click. Autopilot needs Hyperliquid; other exchanges use your own strategy. The video shows setup as simply telling Claude Code to run NOFX, with the panel opening at http://127.0.0.1:3000. First run is guided: fund $1+ USDC on Base for AI fees and $12+ USDC on Hyperliquid to trade. Model choice spans eight providers — DeepSeek, OpenAI, Claude, Qwen, Gemini, Grok, Kimi and MiniMax — or keyless Claw402 (pay-as-you-go in USDC). Philosophy is terse: model proposes, runtime disposes.

Together the trio maps to stages. Freqtrade equals a stable execution engine; you supply the strategy, it runs with discipline, logs everything and stays up. Vibe Trading equals a research and generation lab; it takes a question, backtests, compares via Alpha Zoo and gets it paper-ready. NOFX equals a terminal middle layer; the cable between LLM and exchange with risk clamping and visualization, plus server-side autonomy via Autopilot. That is why the video says try all three: one covers the other's blind spot. A practical chain can be: 1) Invent and prune in Vibe Trading, 2) Harden with a dry run in Freqtrade, 3) Monitor in NOFX and scale with Autopilot on Hyperliquid. Like a kitchen: recipe (Vibe), oven (Freqtrade) and plating (NOFX); without a recipe the oven stays empty, without an oven the recipe stays raw.

Risk is where the video says least and implies most. First, uptime: if the Python terminal closes, orders stop; VPS cost and outage risk are real. Second, API key safety: NOFX claims local encrypted storage, yet any leak equals direct loss; never paste keys into chats and scope permissions to trading only. Third, paper versus live gap: Alpaca paper is a real-time simulation, but slippage (price drift away from the quote), fees and liquidity differ live; both Alpaca and Coinrule docs spell this out. Fourth, backtest overfitting: a strategy that prints 1000% in history can collapse out of sample and in a dry run; Freqtrade docs warn start with dry run before risking funds. Fifth, Hyperliquid perpetual leverage risk: even tokenized equities carry liquidation and funding costs; a heatmap makes the risk visible but not gone. Finally, the small-edge truth seen in Alpha 10132: many academic alphas fade against buy and hold, and a tiny edge can vanish after costs.

Access is the most actionable promise. The narrator says he squeezed installs for all three repos into a single MD file, drag-and-drop into a Claude Code or similar agent session from the Resources section of his free school community. Hostinger VPS discount and Hyperliquid partner links sit in the description. Recommended learning path is clear: paper first, real money later. Invent in Vibe Trading, execute in Freqtrade or NOFX, watch the dashboards, then connect capital gradually. The video notes research agents and trading desks are covered in other episodes and closes with the idea that on the execution side each repo serves a distinct need. Lesson distilled: free execution abundance makes learning by doing cheap; what remains expensive is discipline and patience.