HKUDS/Vibe-Trading

A natural-language finance research agent that gives an AI agent trading-oriented capabilities — multi-market backtesting, factor research, and gated broker connectivity — from a single install.

What it is

Vibe-Trading is a Python agent, installable from PyPI as vibe-trading-ai, that turns natural-language requests into finance research, backtests, and (behind explicit safeguards) live trading actions. It targets developers and quantitatively minded users who want an LLM-driven agent wired to real market data, backtest engines, and broker APIs. The project ships a FastAPI backend, a React web UI, a CLI, and an MCP server, along with a large library of data loaders, backtest engines, and factor definitions.

Key features

Backtesting across multiple markets through dedicated engines (China A-share, China futures, crypto, forex, global equity/futures, India equity, and options portfolios).
A large factor/alpha library (the "Alpha Zoo") plus tooling to register academic factors and monitor their IC/Sharpe decay over a lifecycle.
Many market-data loaders with fallback chains (Yahoo, Binance, OKX, CCXT, Tushare, akshare, Longbridge, MetaTrader 5, and others) and 12 broker connectors.
A multi-agent "swarm" mode and scheduled research, with results deliverable to 16 instant-messaging channel adapters (Telegram, Slack, Discord, Feishu, and more).
An MCP server (supporting Streamable HTTP) exposing tools such as factor analysis and correlation.
Live-trading safety controls described in the changelog: a consent-first mandate gate, kill switch, exposure caps, atomic daily order limits, and an AST-hardened backtest sandbox blocking network/subprocess/eval.

Tech stack

Python >=3.11; project version 0.1.11, MIT-licensed, development status Beta.
Agent layer on LangChain >=1.3.9,<2 and LangGraph >=1.2.5,<1.3 with LangGraph checkpointing.
Backend on FastAPI >=0.104, uvicorn, sse-starlette, pydantic v2, and fastmcp>=2.14; React 19 frontend (per README badges).
Data/quant stack: pandas >=2.0,<3.0, numpy, scipy, scikit-learn, bottleneck, duckdb, ccxt>=4.0, yfinance, akshare, tushare, smartmoneyconcepts.
Reporting/parsing: matplotlib, weasyprint, openpyxl, python-docx, python-pptx, pypdfium2, Pillow.
Optional extras for individual brokers (ibkr, longbridge, mt5), LLM providers (deepseek, anthropic), data sources (ashare/baostock), and each IM channel.

When to reach for it

You want to run backtests and factor research by describing them in natural language instead of hand-writing pipelines.
You need coverage across several markets (A-share, crypto, forex, options, India, global) from one framework.
You want an agent connected to real data loaders and broker read APIs, with research pushed to chat channels.
You want to expose finance tools to other agents over MCP.

When not to reach for it

You want hands-off automated real-money trading; live order paths are deliberately gated behind mandate confirmation and kill switches, and several broker connectors are read-only.
You need a stable, production-grade trading system; the project is pre-1.0 Beta and iterating rapidly.
You want a lightweight single-market backtester; the dependency surface and breadth are substantial.
You are looking for investment advice or guaranteed returns — this is research/engineering tooling, not financial advice.

Maturity signal

The repository has attracted roughly 26k stars since its April 2026 creation and pushes changes almost daily, with a detailed running news log and the last push landing the day before this snapshot — clearly active, heavily maintained development. It is MIT-licensed and not archived, at version 0.1.11 with a Beta classifier and only around 17 open issues, suggesting a small maintainer team keeping pace with a large contributor stream. Read it as young and fast-moving rather than settled; APIs and behavior are still changing.

Alternatives

Use Microsoft Qlib instead when you want an established quant research and factor platform without the LLM-agent layer.
Use Backtrader or Zipline instead when you want conventional, well-documented Python backtesting and don't need the agent, data loaders, or broker breadth.
Use freqtrade instead when your focus is specifically crypto algorithmic trading with a mature bot ecosystem.
Use FinGPT / FinRobot instead when you want LLM-centered financial analysis with a different architecture and scope.

Notes

The README opens with a security warning that a named X account, Virtuals project, and token contract impersonate the project; the maintainers state they have never launched or endorsed any token or memecoin.
Much of the recent changelog is correctness and safety hardening — look-ahead-bias fixes, causal rebalancing, fail-closed live state, and an external security audit whose ten findings were closed — indicating the maintainers treat backtest realism and live-trading safety as first-order concerns.

Wiki: HKUDS/Vibe-Trading