Vibe-Trading - HKUDS Open-Source AI Multi-Agent Financial Workspace

What is Vibe-Trading?

Vibe-Trading is an AI-driven multi-agent financial workspace open-sourced by the University of Hong Kong Data Science Lab (HKUDS), transforming natural language instructions into executable trading strategies, research insights, and portfolio analysis. The project covers global markets including A-shares, Hong Kong stocks/US stocks, cryptocurrencies, futures, and forex, with 6 built-in data sources and automatic fallback mechanisms; most basic data can be used without an API key. The system is equipped with 74 financial professional skills, 29 expert agent teams orchestrated with preset DAGs, and 7 major backtesting engines (including cross-market portfolio backtesting), and supports one-click export to TradingView, Tongdaxin/Tonghuashun, and MetaTrader 5. The project provides four entry points: one-click PyPI installation, zero-configuration Docker deployment, MCP plugin integration, and one-click ClawHub installation. It is compatible with 13+ LLM providers including OpenRouter, DeepSeek, Kimi, and Ollama, and is fully open source under the MIT license.

Main Features of Vibe-Trading

Natural Language Generation of Executable Strategies: Describe trading ideas in plain language, and the AI automatically writes, tests, and exports trading code, covering strategy types such as technical indicators, multi-factor, and machine learning.
74 Financial Professional Skills Library: Divided into 8 major categories (data sources, strategies, analysis, asset classes, cryptocurrencies, capital flows, tools, risk analysis), covering full-chain capabilities from data acquisition to quantitative research.
29 Expert Agent Teams (Swarm): Multi-agent clusters orchestrated with preset DAGs, such as investment committee bull-bear debates, global equity desks, cryptocurrency trading desks, quantitative strategy desks, etc., simulating real institutional collaboration processes.
7 Major Backtesting Engines: Supports A-shares, Hong Kong stocks/US stocks, cryptocurrencies, futures, forex, and cross-market portfolio backtesting, integrating Monte Carlo simulation, Bootstrap confidence intervals, rolling forward validation, and other statistical tests.
One-Click Multi-Platform Strategy Export: Generated strategies can be directly compiled into TradingView Pine Script v6, Tongdaxin/Tonghuashun formulas, and MetaTrader 5 MQL5 code, seamlessly integrating with live trading platforms.
Cross-Session Memory and Skill Evolution: The AI persistently remembers user preferences and historical insights, supports FTS5 session search, and can autonomously create and optimize reusable personal financial skills.
Zero-Configuration Coverage of 6 Major Data Sources: Built-in Tushare, yfinance, OKX, AKShare, CCXT, and Futu, automatically selecting the optimal source with intelligent fallback; most markets can be used without an API key.
Trading Journal and Shadow Account Analysis: Upload broker-exported files (Tonghuashun/Eastmoney/Futu, etc.), automatically analyze trading profiles and behavioral biases, and support shadow backtesting to generate 8-chapter comparison reports.
Universal Intelligent Document Reading: Uniformly parses PDF, Word, Excel, PPT, image OCR, and 40+ text formats, supporting one-click upload and analysis of financial reports, research reports, and trading records.
Real-Time Streaming Interactive Dashboard: The Web UI provides real-time agent status streams, backtest result visualization, interactive charts, and ECharts heatmaps, with front-end support for lazy-loading optimization.
Seamless MCP Ecosystem Integration: Exposes 22 MCP tools (21 without a key), which can be directly integrated into editors such as Claude Desktop, Cursor, and Windsurf, embedding financial capabilities into existing workflows.
Quantitative Analysis Toolkit: Built-in professional modules such as factor IC/IR analysis, Black-Scholes option pricing and Greeks calculation, technical pattern recognition, and MVO/risk parity/BL portfolio optimization.

How to use Vibe-Trading

Option 1: One-click installation via PyPI (recommended)
- Run pip install vibe-trading-ai
- Run vibe-trading init to interactively configure environment variables
- Run vibe-trading to start the interactive TUI, or vibe-trading serve --port 8899 to start the Web service
- Run vibe-trading-mcp to start the MCP service to connect to external agents
Option 2: Zero-configuration deployment with Docker
- git clone repository, copy agent/.env.example to agent/.env and fill in the LLM API Key
- Run docker compose up --build
- Open http://localhost:8899 in your browser to use it
Option 3: Integrate MCP plugin into an existing Agent
- In the Claude Desktop configuration, add the vibe-trading-mcp command
- Or configure stdio/SSE transport in OpenClaw, Cursor, or Windsurf
- 21/22 tools can be called without any API Key
Option 4: One-click installation via ClawHub
- Run npx clawhub@latest install vibe-trading --force
- Automatically downloads the skill package and MCP configuration to the Agent directory, with no manual cloning required

The official website address of Vibe-Trading

Github repository：https://github.com/HKUDS/Vibe-Trading

Key information and usage requirements for Vibe-Trading

Python version requirement：3.11+。
PyPI package name：vibe-trading-ai(the command after installation is vibe-trading、vibe-trading serve、vibe-trading-mcp）。
LLM configuration: needs to be configured in agent/.env configure LANGCHAIN_PROVIDER and the corresponding API Key; supports Ollama local key-free operation.
Recommended models: For complex scenarios, Claude Opus/GPT-5.4/Gemini 3.1 Pro are recommended; for everyday use, cost-effective models such as DeepSeek-V3.2, Grok-4.20, and Kimi K2.5 are recommended.
Data sources: Tushare Token is optional (AKShare free fallback); yfinance (Hong Kong stocks/US stocks), OKX (cryptocurrencies), and CCXT (100+ exchanges) are all key-free.
Web UI ports: development mode frontend 5899(proxied to the backend), production mode single service 8899。
File upload limit: Web UI streaming upload, single file limit 1MB, to prevent memory overflow.
Open-source license：MIT License。
Latest version: v0.1.6 (released on 2026-04-28).

Core advantages of Vibe-Trading

Zero-threshold global market coverage: Integrated coverage of six major markets—A-shares, Hong Kong stocks, US stocks, cryptocurrencies, futures, and forex—with most basic data free and requiring no Key.
Agent cluster collaboration: Not single-Agent Q&A, but 29 preset teams collaborating through DAG orchestration, simulating real investment research/trading/risk control processes.
Strategies can be implemented and executed: The generated strategies are not demo code; they can be directly compiled to live trading platforms such as TradingView, Tongdaxin, and MT5.
Continuous evolution across sessions: Equipped with a persistent memory layer, every conversation trains your dedicated trading assistant instead of repeatedly starting from scratch.
Statistical rigor: Backtest results come with Monte Carlo simulation, Bootstrap confidence intervals, and rolling forward validation to avoid overfitting traps.
Seamless MCP ecosystem integration: 22 MCP tools (21 of which require no key) can be directly integrated into editors such as Claude Desktop, Cursor, and Windsurf.
Open source and extensible: MIT license, Python 3.11+, supports zero-cost local operation with Ollama, and can also connect to commercial APIs such as OpenRouter.

Comparison of Vibe-Trading with similar competitors

Comparison dimension	Vibe-Trading (HKUDS)	Composer.trade	TradingAgents (TauricResearch)
Product positioning	AI multi-agent financial workspace, natural language → executable strategies + cross-market backtesting + multi-platform export	AI-native no-code investment platform, natural language → automated trading strategies (Symphony), vertically integrated trading ecosystem	Multi-agent LLM financial trading framework that simulates the collaborative decision-making process of a real trading firm (analyst → researcher → trader → risk control)
Open-source license	Fully open source under MIT, supports local deployment and secondary development	Closed-source commercial product, core strategy logic is not open	Open-source academic research framework, built on LangGraph, extensible code
Agent architecture	29 preset DAG-orchestrated Swarm teams(investment committee, global equity desk, cryptocurrency desk, etc.), supports streaming status monitoring	Single-Agent strategy generation + community strategy cloning, no concept of multi-agent collaboration	4-layer multi-agent: analyst team (fundamentals/sentiment/news/technical) → research team (bull-bear debate) → trader → risk control/PM final decision
Natural language interaction	Supported, after describing the idea the AI automatically writes code, backtests, and exports to third-party platforms	Supported, converts natural language into executable, automatically tradable strategies within 60 seconds	Supported, enter the ticker and date via CLI, and the Agent automatically completes the analysis and outputs trading decisions
Market coverage	A-shares, Hong Kong stocks, US stocks, cryptocurrencies, futures, forex, six major markets, supports cross-market portfolio backtesting	US stocks, ETFs, cryptocurrencies, options, limited to the US market	PrimarilyUS stocks(data obtained via yfinance/Alpha Vantage)
Data sources and configuration	6 major data sources（Tushare、yfinance、OKX、AKShare、CCXT、Futu），automatic fallback, API-key-free for most markets	Proprietary data infrastructure, trade execution integrated through the Alpaca broker	yfinance, Alpha Vantage, FinnHub; local Tauric TradingDB (in development)
Backtesting capability	7 major backtesting engines(including a cross-market portfolio engine), integrating Monte Carlo, Bootstrap confidence intervals, rolling forward validation, and 4 types of optimizers	Sub-second backtesting, historical performance validation, AI overfitting detection	Supports backtesting, based on the self-built TradingDB dataset (not yet fully released)
Strategy export and live trading	One-click export to multiple platforms: TradingView Pine v6, Tongdaxin/Tonghuashun, MetaTrader 5 MQL5; no direct broker connection	Automated execution within the platform, direct trading and rebalancing through your own brokerage account; export to external platforms not supported	Connects only tosimulated exchangesfor execution, does not support live trading via real brokerage APIs
MCP plugin support	22 MCP tools(21 key-free), can connect to Claude Desktop, Cursor, Windsurf and other editors	Officially provides composer-trade-mcp, supports strategy creation and backtesting in Claude/Cursor	No native MCP support
Cross-session memory	Persistent memory + FTS5 search + self-evolving skills(Agent can create and optimize reusable skills across sessions)	No cross-session memory, each strategy creation is an independent task	No cross-session memory, state is not retained after a single analysis workflow ends
Quantitative professional tools	Factor IC/IR analysis, Black-Scholes Greeks calculation, technical pattern recognition, MVO/risk parity/BL portfolio optimization	Basic technical indicators (RSI/MA/EMA, etc.), AI-assisted parameter tuning	Technical indicators (MACD/RSI), social media sentiment scoring, news macro analysis, fundamental financial metrics
Community and ecosystem	One-click installation via ClawHub, OpenSpace self-evolving skill sharing, 74 extensible financial skills	3000+ community strategy marketplace, can clone and adapt others' strategies	Academic research-oriented, Tauric Research community, encourages academic citation
Pricing model	Completely free(only requires your own LLM API Key, supports zero-cost local operation with Ollama)	Freemium: free (backtesting only); $30/month (1 live strategy); $120/month Pro (unlimited strategies)	Free and open source(requires your own LLM and data API Key)
Target audience	Individual quantitative traders, A-share/Hong Kong stock/cryptocurrency investors, AI Agent developers, finance students	US retail investors, ordinary users with no programming background who want to automate trading	Academic researchers, AI finance experimenters, US stock traders who want to simulate institutional decision-making processes

Vibe-Trading use cases

24/7 cryptocurrency monitoring: Analyzes perpetual contract funding rates and basis, liquidation heatmaps, stablecoin capital flows, DeFi yields and on-chain whale movements, providing multi-dimensional data support for cryptocurrency trading.
Intelligent parsing of earnings reports and research reports: Upload PDF earnings reports, Word research reports, Excel financial models or PPT roadshow materials, and the system automatically extracts key risk points, earnings beat/miss items, and changes in management guidance, generating a structured investment memorandum.
AI Agent financial capability plugin: Through 22 MCP tools (21 key-free), Vibe-Trading can be connected to editors such as Claude Desktop, Cursor, and Windsurf, giving general-purpose AI assistants professional financial skills such as real-time quote queries, backtest execution, and strategy generation.
Localized privacy-compliant deployment: Achieve full-chain private operation through local large models via Docker or Ollama. Trading data and strategy logic do not leave the local environment, meeting the strict requirements of financial institutions or high-net-worth individuals for data sovereignty and compliance.
A-share risk pre-screening: Use the ashare-pre-st-filter skill to automatically screen for A-share ST/*ST risks and administrative penalty relevance, and combine with the Sina data source to filter out securities account-related noise, thereby avoiding delisting and regulatory risks in advance.