Vibe-Trading is an open-source finance research workspace that turns natural-language questions into runnable analysis. A prompt is routed to finance skills, market-data loaders, generated strategy code, and a matching backtest engine, and the run returns reports, run cards, tool traces, and exports that later sessions can reuse. It is built for research, simulation, and backtesting rather than brokerage execution: live trading is opt-in, read-only by default, and runs only through a broker the user authorizes, within limits the user sets. It holds no funds and runs no execution venue.
Each run follows one evidence path. Plan selects the relevant skills, tools, data sources, and an optional swarm preset. Ground loads equities, crypto, futures, forex, documents, or web context through the available loaders. Execute generates testable strategy code and runs the matching backtest engine. Validate adds metrics, benchmark comparison, Monte Carlo, bootstrap, walk-forward, run cards, and warnings. Deliver returns reports, artifacts, tool traces, and exports for TradingView Pine Script, TDX, MetaTrader 5, and MCP clients. A no-progress guard stops the agent after eight consecutive tool calls without a new successful observation and asks for a recovery decision instead of looping.
Market data comes from 28 sources behind one call. With automatic source selection, the loader picks a per-market fallback chain ordered by IP-ban risk, with keyless public sources first and throttled or key-gated ones last, so A-shares, HK, US, Canada, UK, India, Korea, Vietnam, crypto, futures, and forex all work without an API key. Beyond price bars, 22 read-only data tools cover fund flow, SEC filings, financial statements, options chains, 13F holdings, ETF look-through, prediction markets, and research papers. An optional QVeris marketplace adds paid premium data on an explicit-only basis and never joins the automatic fallback chain.
The agent runs as an interactive terminal UI, as a REST and SSE API server with a web UI, as an MCP server exposing 76 tools to clients such as Claude Desktop, OpenClaw, and Cursor, and as IM channel adapters (Telegram, Slack, Discord, Matrix, WhatsApp, Signal, Feishu, DingTalk, Teams, email, and others) that share the same session runtime. It is built on LangChain and selects its LLM provider by configuration: OpenRouter, OpenAI, Anthropic, DeepSeek, OpenCode (Go and Zen), Gemini, Groq, DashScope/Qwen, Zhipu, Moonshot/Kimi, MiniMax, SiliconFlow, Xiaomi MIMO, Novita AI, iFlytek Spark, Z.ai, NVIDIA NIM, ModelScope, GitHub Copilot, and Ollama for local models. OpenAI Codex can also be used through ChatGPT OAuth, and GitHub Copilot through a Copilot subscription. Because skills, backtests, memory, and swarms all flow through tool calls, model choice decides whether the agent uses its tools or fabricates answers.
| Tier | Models |
|---|---|
| Best | Claude Opus 4.7, Claude Sonnet 4.6, GPT-5.5 Pro, Gemini 3.5 Flash (complex swarms of three or more agents, long research sessions) |
| Default | DeepSeek V4 Pro, Grok 4.20, GLM-5.1, Kimi K2.6, Qwen3 Max Thinking (daily use, reliable tool calling at about a tenth of the cost) |
| Not recommended for agent use | nano, flash-lite, coder-next, and other small or distilled variants (tool calling is unreliable, so the agent answers from training data instead of running tools) |
Features
- Natural-language research: turns a market question into routed skills, fetched data, generated strategy code, a backtest, and a run card, with persistent memory and editable skills that make repeated routines reusable
- Backtest engines: ten engines plus an options portfolio engine: China A-share (T+1, price limits, pre-ST filter), global equity, India, Korea, Vietnam, crypto spot and USD-M perps, China and global futures, forex, and a cross-market composite with one shared capital pool
- Shadow Account: parses broker exports, profiles behavior (holding days, win rate, disposition effect, overtrading, momentum chasing, anchoring), extracts recurring rules into a strategy profile, backtests them, and renders an HTML or PDF audit report
- Multi-agent swarms: 30 preset research teams (investment committee, quant desk, risk review, macro, crypto) with streaming progress, persisted reports, workers grounded in fetched market data, and operator-approved tools from external MCP servers
- Alpha Zoo: 462 pre-built cross-sectional alphas (Qlib 158, Kakushadze 101, GTJA 191, academic, PIT-safe fundamental) benched in one command with IC, IR, and alive/reversed/dead categorization
- Finance skill library: 90 skills in nine categories covering data sources, strategy, analysis, asset classes, crypto, flow, tools, research, and risk
- Data sources with fallback: 28 sources behind one call, per-market chains ordered by IP-ban risk, local CSV, Parquet, or DuckDB files, and a hook for registering a custom historical-bar loader
- Read-only research tools: 22 data tools for fund flow, dragon-tiger, northbound, margin, block trades, SEC filings, financial statements, options chains, 13F holdings, ETF look-through, prediction markets, and research papers
- Broker connectors: 18 brokers (IBKR, Robinhood, Alpaca, Tiger, Futu, OKX, Binance, eToro, MetaTrader 5, KIS, and others) with read, paper, and bounded-live modes, a read-only aggregated Portfolio page, and a TAP mode with isolated credentials and human-approved writes
- Document and chart reading: parses PDF, DOCX, XLSX, PPTX, and images with pluggable OCR, and reads chart screenshots with a vision model
- MCP server and client: exposes 76 MCP tools as a stdio subprocess for Claude Desktop, OpenClaw, Cursor, and other clients, and loads tools from external MCP servers in turn
- Delivery surfaces: terminal UI, REST and SSE API with a web UI, scheduled research jobs on an interval or cron, and IM channels including Telegram, Slack, Discord, Matrix, WhatsApp, Signal, Feishu, DingTalk, Teams, and email
