What is Vibe-Trading?
Vibe-Trading is an open-source finance research agent from the Data Intelligence Lab at the University of Hong Kong (HKUDS, also behind LightRAG). You ask a market question or describe a strategy in plain language; it fetches data, writes strategy code, runs a backtest and returns metrics, reports and exportable code for TradingView Pine Script, TDX or MetaTrader 5. It runs as a CLI, a local web UI, a REST API, an MCP server with 84 tools, and through chat channels from Telegram and Slack to WeChat and Feishu.
It covers equities in China, Hong Kong, the US and several other markets, plus crypto, futures and forex, from 29 built-in free data sources with fallback; a paid QVeris marketplace is opt-in. A Shadow Account feature reads a broker trade export, extracts the rules you actually trade by and backtests them against what you did. Broker connectors default to read-only; live orders through Robinhood, Alpaca, Binance, eToro and others need a mandate you set, with a kill switch.
The repository was created in April 2026, is MIT-licensed with about 37,300 GitHub stars, and is at v0.1.18 (21 October 2026) on PyPI as vibe-trading-ai. The README warns that a token and X account using its name are not official.
Is Vibe-Trading worth it?
Vibe-Trading is more research workstation than trading bot, and that is its strength. The core loop turns a sentence such as "backtest a BTC-USDT 20/50 moving-average strategy for 2024" into strategy code, a backtest with benchmark comparison, validation artefacts and a run card that records how the result was produced. A bundled alpha zoo benches 478 published factors on your universe in one command, and Monte Carlo, bootstrap and walk-forward validation are built in.
The changelog is the most persuasive document in the repository. October's entries are almost all financial-correctness fixes: Sortino downside deviation, first-bar PnL in risk ratios, SEC filing timing in factor tests, A-share board limits, settlement-currency accounting. That is what a project looks like when users run real numbers through it, and it is also a warning that results from earlier versions may have been wrong in ways those fixes describe.
Security defaults are sensible for a local tool. Docker publishes on 127.0.0.1:8899 as a non-root user, remote API access is refused without an API_AUTH_KEY, shell tools are off unless enabled, and live trading requires a mandate of symbols, size, exposure and daily cap plus a filesystem kill switch. The README is candid that its tool-prefix filters are "not a security boundary" because generated Python can still run arbitrary code.
The breadth is also the risk. It is seven months old, at v0.1.18, with about 50 broker profiles across paper and live modes, many described as declared rather than verified, and quality depends heavily on the model: the docs warn that small and flash-lite models fabricate answers instead of calling tools. Compared with TradingAgents, which produces a debated buy-or-sell call on one ticker, Vibe-Trading is the better tool for testing a strategy and the worse one for a quick, reproducible opinion.
Where it earns its place
A retail trader with two years of broker exports
See which of their own habits cost them money
Upload the CSV and ask for a Shadow Account; the agent extracts the trader's rules, backtests them and writes an eight-section report on early exits, rule breaks and missed signals. Free apart from LLM tokens.
A quant researcher at a small Hong Kong fund
Screen published factors on CSI 300 before building anything new
Run vibe-trading alpha bench against the GTJA 191 or Qlib 158 zoo for 2018 to 2025 and read the alive, reversed and dead split. It replaces a week of factor plumbing, though results should be re-checked in the fund's own stack.
A developer building a finance assistant in Claude Desktop
Give the assistant real market data and backtests as tools
Add Vibe-Trading's stdio MCP server; core research tools need no API keys for HK, US and crypto data. Keep broker profiles read-only unless a mandate is configured deliberately.
Before you commit
- Setup
- pip install vibe-trading-ai on Python 3.11+, or docker compose up --build; run vibe-trading init for the .env, add one LLM key, then vibe-trading run -p with a prompt. The web UI is at localhost:8899.
- Integrations
- About 20 LLM providers including OpenAI, Anthropic, DeepSeek, Gemini, OpenRouter and local Ollama; 29 free data sources plus key-gated Finnhub, Alpha Vantage, Tiingo and FMP; broker SDKs from Alpaca and IBKR to Futu and Binance; MCP and more than a dozen chat channels.
- Data and privacy
- Runs locally; memory, sessions, backtests and broker config sit in local files or named Docker volumes. Prompts go to your chosen LLM provider and data requests to the market-data sources. Remote API access requires API_AUTH_KEY. MIT licence.
- Cost at scale
- No paid edition. Costs are LLM tokens per research run (swarm runs use several agents) and any premium data keys; the docs advise against nano and flash-lite models, which pushes users towards pricier ones.
- Limits
- About 22 open issues and pull requests; release notes flag many fixes as unreleased on main; native Windows confirmation for paper strategies is pending; Robinhood options execution is unavailable.
What’s good
- Natural-language backtests with validation and run cards
- 478-factor alpha zoo and cross-market composite backtests
- 84 MCP tools; CLI, web UI, API and chat channels
- Live orders off by default and bound by a mandate and kill switch
What’s not
- v0.1.x, seven months old, with frequent correctness fixes
- Many broker profiles are declared, not verified
- Small models skip tools and invent results
- Generated code runs locally; filters are not a sandbox
Best for
Retail quants and finance researchers who want to test strategy ideas in plain language and audit how each result was produced.
Vibe-Trading alternatives
Vibe-Trading: common questions
- Is Vibe-Trading worth it?
- Vibe-Trading packs a quant desk's worth of backtesting and data plumbing behind a chat box, and its unusually careful changelog is the best reason to trust it more than the average trading bot. We rate it 3.9 out of 5.
- How much does Vibe-Trading cost?
- Free and open source (MIT). You pay your LLM provider; market data is free by default, with optional paid sources (QVeris, Finnhub, Tiingo and others) on your own keys.
- What are the downsides of Vibe-Trading?
- v0.1.x, seven months old, with frequent correctness fixes; many broker profiles are declared, not verified; small models skip tools and invent results; generated code runs locally; filters are not a sandbox.
- Who is Vibe-Trading best for?
- Retail quants and finance researchers who want to test strategy ideas in plain language and audit how each result was produced.
- How much does Vibe-Trading cost for a team or in production?
- No paid edition. Costs are LLM tokens per research run (swarm runs use several agents) and any premium data keys; the docs advise against nano and flash-lite models, which pushes users towards pricier ones.
- Is Vibe-Trading safe for company data?
- Runs locally; memory, sessions, backtests and broker config sit in local files or named Docker volumes. Prompts go to your chosen LLM provider and data requests to the market-data sources. Remote API access requires API_AUTH_KEY. MIT licence.
- What are the best Vibe-Trading alternatives?
- The closest alternatives we have reviewed are TradingAgents.
Answers checked Oct 31, 2026.