Overview
Vibe-Trading is an open-source research workspace from HKUDS that connects natural-language prompts to market-data loaders, strategy generation, backtest engines, reports, exports and persistent research memory. You ask a finance question in plain language and it answers by running real tools against real data, leaving behind strategy code, metrics, validation artifacts and run cards rather than a chat transcript.
The project is built for research, simulation and backtesting. It can also trade autonomously through a broker you authorise yourself, but it holds no funds, never trades outside the limits you set, and can be halted instantly. Broker credentials go to the OS keyring rather than config files, and every run writes a hash manifest over the prompt, skills, tool registry and package versions plus a hash-chained audit ledger, so you can answer "what methodology produced that number?" after the fact.
Beyond the single agent, Vibe-Trading runs multi-agent analyst teams — investment, quant, crypto, macro and risk workflows — with streaming progress and persisted reports. It reaches markets across A-share, Hong Kong, US, Canada, UK, India and Korea equities plus crypto, futures and forex, ships a pre-built alpha zoo you can bench in one line, and exposes the same runtime through a CLI, a FastAPI web UI, a REST API and an MCP server that plugs into Claude Desktop, Cursor and other MCP clients.
What it does
- Natural-language market research with tool calls, document and chart reading, and reusable session context
- Multi-agent trading teams for investment, quant, crypto, macro and risk reviews, with streaming progress and persisted reports
- Cross-market backtesting across A-share, HK, US, Canada, UK, India and Korea equities plus crypto, futures and forex, with point-in-time data, validation artifacts and run cards
- A pre-built alpha zoo benched in one command, categorising alphas as alive, reversed or dead on your own universe
- Broker connectors kept read-only or paper by default, with credentials in the OS keyring, a committed mandate, an order guard and an instant halt
- Governance in every run: a hash manifest over prompt, skills, tool registry and package versions, plus a hash-chained audit ledger
- One runtime behind four transports — interactive CLI/TUI, FastAPI web UI, REST API, and an MCP server for Claude Desktop, Cursor and other clients
- Exportable artifacts including reports, TradingView Pine Script, TDX and MetaTrader 5
Getting started
Vibe-Trading is on PyPI as `vibe-trading-ai` and needs Python 3.11+ and an LLM API key from any supported provider — or a local Ollama install, which needs no key.
Install from PyPI
One pip install gives you three commands: `vibe-trading` (CLI/TUI), `vibe-trading serve` (web server) and `vibe-trading-mcp` (MCP server).
pip install vibe-trading-aiConfigure a provider
`init` walks you through an interactive .env setup. Supported providers include OpenRouter, OpenAI, Anthropic, DeepSeek, Gemini, Groq, DashScope/Qwen, Zhipu, Moonshot/Kimi, MiniMax and local Ollama.
vibe-trading initRun a first research task
Pass a prompt with -p and the agent fetches data, writes and runs the backtest, and summarises the result.
vibe-trading run -p "Backtest a BTC-USDT 20/50 moving-average strategy for 2024 and summarize return and drawdown"Bench the alpha zoo
Score a whole family of pre-built quant alphas against your universe and period in one line.
vibe-trading alpha bench --zoo gtja191 --universe csi300 --period 2018-2025 --top 20Launch the web UI or the MCP server
`serve` starts the FastAPI web UI; `vibe-trading-mcp` starts the MCP server over stdio so Claude Desktop, Cursor or another MCP client can call the same tools.
vibe-trading serve --port 8899
vibe-trading-mcpCommands and code are distilled from the project's own documentation — always check the official repo for the latest.
When to use it
- Ask a market question in plain language and get an answer backed by fetched data, tool calls and a reproducible run record
- Prototype and backtest a strategy idea across several markets without wiring up data loaders and an engine yourself
- Run a team of analyst agents over the same thesis to get investment, quant and risk views side by side
- Give an existing MCP client — Claude Desktop, Cursor, OpenClaw — a set of read-only market-data and quant tools
How Vibe-Trading compares
Vibe-Trading alongside other open-source multi-agent systems tools AI/TLDR tracks, ranked by GitHub stars.
| Tool | Stars | What it does |
|---|---|---|
| TradingAgents | ★ 105k | LangGraph framework that mirrors a trading firm: LLM analyst, bull/bear researcher, trader and risk-management agents debate before a decision. For research, not advice. |
| Ruflo | ★ 72k | Agent meta-harness that wraps Claude Code and Codex with 100+ specialized agents, swarm coordination, vector memory, background workers and cross-machine agent federation. |
| MetaGPT | ★ 70.3k | A multi-agent framework that models a software company, assigning roles like product manager, architect, and engineer to generate code from a single prompt. |
| AutoGen | ★ 60.9k | Microsoft Research's framework for building applications where multiple agents converse with each other and with tools to solve tasks. |
| CrewAI | ★ 58.4k | A framework for assembling teams ('crews') of role-playing agents that divide tasks and collaborate to complete a goal. |
| Vibe-Trading | ★ 33.2k | An open-source research workspace that turns finance questions into runnable analysis, backtests and multi-agent reviews |
| AgentScope | ★ 31.3k | A framework for building multi-agent applications with message passing, visual debugging tools, and distributed execution. |
| OpenAI Agents SDK | ★ 29.3k | OpenAI's lightweight Python SDK for building multi-agent workflows using explicit handoffs, tools, and guardrails. |