Vibe-Trading

• Backtesting a moving-average or RSI strategy on BTC-USDT, US equities, or A-shares from a single natural-language prompt
• Running an investment-committee swarm on a stock to get bull, bear, and risk views before a decision
• Benchmarking the GTJA 191 or Alpha101 factor set on the CSI 300 over several years
• Uploading a broker trade export to see behavioral patterns and how a rule-based version of your own trading would have done
• Scheduling a pre-market brief or portfolio check-up that is delivered to Telegram, Feishu, or email
• Giving Claude Desktop or Cursor finance research and backtest tools through the MCP server

Vibe-Trading is less a single trading bot than a full research bench that an LLM drives: data routing across many markets, market-aware backtest engines, factor benchmarking, and a strict stance on not inventing numbers or inputs. That makes it a strong fit for people who want to test ideas quickly and inspect how each figure was produced. The tradeoffs are breadth and pace. There is a lot to configure, the changelog shows a steady stream of calculation fixes, and live trading is opt-in, capped, and explicitly experimental. Treat it as a research and paper-trading tool, pin versions when results matter, and keep any live mandate tightly capped if you enable one.

About Vibe-Trading

Vibe-Trading is a self-hosted research workspace published under the HKUDS GitHub organization, whose profile names it the Data Intelligence Lab@HKU. You install it from PyPI (`pip install vibe-trading-ai`), from source, or with Docker, then work through an interactive terminal UI, single-shot CLI runs, a local React web UI served by a FastAPI backend, a REST API, IM channels such as Telegram, Slack, Discord, Feishu, and email, or an MCP server that exposes 76 tools to Claude Desktop, Cursor, OpenClaw, and other MCP clients. A request is routed to the relevant finance skills (90 ship with it) and data sources, grounded in fetched market data, executed as generated strategy code or analysis tools, validated with metrics, benchmark comparison, Monte Carlo, bootstrap, and walk-forward checks, and delivered as a report, run card, or Pine Script, TDX, or MetaTrader 5 export. Market coverage spans A-shares, Hong Kong, US, Canadian, UK, Indian, Korean, and Vietnamese equities, crypto spot and perpetuals, futures, and forex, through 28 market-data sources with automatic fallback, and the free sources need no API keys. Backtests run on market-specific engines that model rules such as A-share T+1 and price limits, Indian circuit bands, and crypto funding, plus a cross-market composite and an options-portfolio engine. Beyond single-agent chat it offers 30 preset multi-agent teams (an investment committee, a quant desk, a crypto desk, a risk committee), a 462-factor Alpha Zoo you can bench on your own universe, a tested quant library, DCF and comps models that refuse to run on missing inputs, persistent memory with cross-session search, scheduled research jobs, and a Shadow Account that parses your broker exports, extracts your implicit trading rules, and backtests them against what you actually did. Eighteen broker connectors cover IBKR, Robinhood, Alpaca, Tiger, Futu, OKX, Binance, eToro, MetaTrader 5, KIS, Zerodha, and others. Most offer read access and paper orders; live order placement exists only on some of them and only inside a mandate you set (symbol allowlist, order-size, exposure, and daily trade caps) with an instant kill switch, and order-placing tools are never exposed over MCP. The project says it is research software, not investment advice, that its broker-trading capability is experimental and has not been verified by the project against a real broker account, and that past performance does not guarantee future results. It is free to use; you pay your own LLM provider (or run a local model through Ollama) and any paid data feeds. The repository had about 34.8k GitHub stars on October 5, 2026, and ships near-daily changes, including frequent fixes to metric and factor calculations.