Vibe-Trading

What it is

Vibe-Trading is a finance research agent you install and run on your own machine, published by HKUDS — the University of Hong Kong Data Intelligence Lab. pip install vibe-trading-ai, give it one LLM provider key, and ask in plain language: it loads from 90 bundled skills, pulls bars from whichever of its 27 data sources answers first, writes and runs strategy code, and returns a run card with metrics and a tool trace. The same install ships a CLI, a FastAPI web UI, and an MCP server exposing 74 tools over stdio or streamable HTTP — which is how most people attach it to an agent they already use.

It is new. First commit 1 April 2026, v0.1.15 released 9 September 2026, last commit to main on 13 September 2026, MIT-licensed, not archived, roughly 33,300 stars. The star count runs a long way ahead of the shipping record.

Pricing

Nothing to buy: no hosted tier, no paid build, no data bundle. The cost is the model you point it at, plus optional keys for premium data — including QVeris, a third-party credit-metered marketplace the README links behind a referral code.

Data & coverage

It ships loaders, not a feed. One get_market_data call fans out across 27 sources with a fallback chain ordered by IP-ban risk, and most need no key. The loader layer is point-in-time historical bars only — the project states that real-time ticks and order-book depth are out of scope and come from broker connectors instead. Equities cover A-shares, Hong Kong, US, Canada, UK, India and Korea; crypto via OKX, Binance and CCXT; forex via MetaTrader 5. Nine backtest engines plus a cross-market composite and an options-portfolio engine sit behind them, with Monte Carlo permutation, bootstrap Sharpe intervals and walk-forward validation.

Integrations

Fourteen broker connectors, read-only or paper by default. Strategies export to TradingView Pine Script, TDX, MQL5 and vn.py; sessions as Markdown or JSON; portfolio snapshots as CSV. It also loads tools from external MCP servers, so it is a client as well as a server.

Limitations

• No LLM of its own and no free inference. A weak model degrades it silently into answering from training data instead of running the tools.
• Live trading is the project's own "experimental and not verified by us against a real broker account". IBKR is read-only; Trading 212 refuses orders even on paper.
• The free data sources are public endpoints nobody promised to keep. A September 2026 fix documents eight years of requested history silently returning as two.
• The Electron desktop shell is built from source with Windows-only packaging; releases publish no installer.
• Version 0.1.x, five months old, and upgrades have broken once already — 0.1.10 moved to LangChain 1.x and needs a fresh virtualenv.
• It emits opinions on named securities: the shipped investment_committee preset ends in a "PM final call", which is a language model's view, not research you can audit.

Alternatives

For read-only market data over MCP without an agent framework attached, single-purpose servers — SEC EDGAR MCP, or a data vendor's own MCP endpoint — are far less machinery. For backtesting without an LLM in the loop, NautilusTrader, QuantConnect LEAN and vectorbt do that job directly. AI-Trader, from the same lab, is a hosted platform rather than a local workspace.