quantvibe

QuantVibe is an integration project that connects Microsoft Qlib and HKUDS Vibe-Trading without forking either tool. Qlib trains machine learning models on market data and produces stock rank scores, while Vibe-Trading consumes those scores through a read-only FastMCP server and acts on them, with paper trading as the default execution mode. The bridge is implemented in roughly 600 lines of Python with no heavy dependencies. The two systems never import each other; they communicate through an immutable signed contract (artifacts/signals.json with SHA-256 checksums) and a FastMCP stdio server. Each side runs in its own virtual environment to avoid dependency conflicts. Key features include a one-command demo mode using synthetic trend data and a momentum fallback, a real mode that downloads OHLCV data via yfinance, converts it to Qlib binary format, computes Alpha158 features, and trains a LightGBM ranker. Before signals are published, an evaluation gate checks Spearman correlation (IC), ICIR, and top-k hit-rate against forward returns. A SQLite track record ledger settles historical signals against realized prices and reports hit-rate and excess return. The FastMCP server exposes three read-only tools: get_latest_signals, list_universe, and signal_health. Execution is double-guarded: real broker order submission requires both the --submit flag and the VIBE_ALLOW_ORDERS=1 environment variable. Additional connectors include a MetaTrader 5 native EA bridge for Forex, CFDs, and prop firms, plus a Bloomberg-style RSS financial news stream with multi-universe monitoring. The project includes a modern fintech web terminal built with FastAPI, React 19, TypeScript, and Tailwind CSS. It provides a dashboard with model quality gates, an interactive pipeline launcher with SSE streaming, an execution desk with paper/live safety switch, track record audit views, and an MCP inspector. The frontend source lives in web/frontend/ and compiles to static assets in web/static/. Deployment options include isolated virtual environments via scripts/setup.ps1, Docker with a prebuilt image on GitHub Container Registry, and a 24/7 background service using nohup. The project is distributed under the MIT License and includes a legal notice emphasizing educational and research use only.