Vibe-Trading is an open-source research workspace that turns finance questions asked in plain language into runnable analysis. A prompt is routed to market-data loaders, strategy generation, a backtest engine, validation, and report export, with persistent memory and editable skills so a useful routine becomes a reusable workflow. It is built for research, simulation, and backtesting, and will trade only through a broker you authorize yourself, within limits you set and with an instant halt.
Market data comes from 23 free sources behind one getmarketdata call. With source: "auto" the loader picks by symbol and walks a per-market fallback chain ordered by IP-ban risk, putting never-banned public sources first and throttled or key-gated ones last. Coverage spans A-share, Hong Kong, US, Canadian, UK, Indian, and Korean equities, crypto through OKX, CCXT, and Binance, futures, forex and metals through MetaTrader 5, and your own CSV, Parquet, or DuckDB files under a local: prefix. Every served frame states its price caliber — adjusted, raw, or unknown — and a mixed-caliber run says so. A further 22 read-only tools reach fundamentals and flow data, SEC filings, 13F holdings, ETF constituents, options chains, macro series, prediction markets, and research papers.
Backtesting produces inspectable artifacts rather than a single number: generated strategy code, metrics, benchmark comparison, Monte Carlo, bootstrap and walk-forward validation, and run cards. Shadow Account works from your own broker export instead of a template — it parses the journal, profiles behaviour with holding days, win rate, drawdown, disposition effect, overtrading and anchoring checks, extracts your recurring entry and exit rules into an explicit strategy, backtests that shadow, and reports rule breaks, early exits, and missed signals. An alpha zoo of 462 pre-built alphas benches in one line, and multi-agent swarms run investment, quant, crypto, macro, and risk reviews.
The same runtime is reachable several ways: an interactive terminal interface, a FastAPI server with a React web interface, a stdio MCP server exposing 74 tools to any MCP client, and messaging channels including Telegram, Slack, Discord, Matrix, and email. A read-only portfolio page aggregates holdings across broker connections that declare only account and position read scopes, valuing them in US dollars with per-source provenance and excluding any source that errors. Broker credentials are collected in a local terminal prompt or connection form and stored in the operating-system keyring, never in a prompt or an MCP call.
Model choice is left to you across roughly twenty providers, from hosted APIs to a local Ollama install that needs no key, configured through LANGCHAIN_PROVIDER, a provider key, a base URL, and a model name. Because the agent is tool-heavy, the project groups models into tiers and notes that small, distilled, or -lite variants call tools unreliably and tend to answer from training data instead of loading skills or running backtests.
Features
- Natural-language research: a prompt is routed through data loading, tool execution, validation, and report delivery, with session memory
- 23 free data sources: per-market fallback chains ordered by ban risk cover A-share, HK, US, Canada, UK, India, Korea, crypto, futures, and forex
- Price caliber reporting: each frame declares adjusted, raw, or unknown pricing, and mixed-caliber runs are flagged
- Backtest validation: metrics, benchmark context, Monte Carlo, bootstrap, walk-forward tests, and reproducible run cards
- Shadow Account: broker journals are parsed, behaviour is profiled, rules are extracted, and the shadow strategy is compared with actual trades
- Alpha zoo: 462 pre-built alphas from Qlib, Kakushadze, GTJA, academic, and fundamental sets bench in a single command
- Multi-agent swarms: investment, quant, crypto, macro, and risk teams run with streaming progress and persisted reports
