Inspired by TauricResearch/TradingAgents — re-implemented for China A-share markets as a native desktop app with Tauri + Vue 3, a provider-agnostic LLM layer, and real-time streaming UI.
A multi-agent stock analysis and strategy backtesting desktop app for China A-share markets.
Analysis: Pick one or more stocks, set a date range, and watch eight specialised agents work in sequence — results stream to the UI in real time.
Backtesting: Define trading strategies via preset templates or natural language (translated by LLM), and replay them over historical data with a deterministic engine that enforces A-share rules (T+1 settlement, 100-share lots, commission + stamp tax).
User picks stock + date range
│
▼
┌─────────────────────────────────────────────────────┐
│ 1. Chart Candlestick + MACD + RSI (Plotly)│
│ 2. Market Technical analysis (tools) │
│ 3. Fundamental Financials & valuation (tools) │ ← parallel
│ 4. News Sentiment from recent articles │
│ 5. Bull Pro-buy argument (streaming) │ ← waterfall
│ 6. Bear Pro-sell argument (streaming) │
│ 7. Trader Investment decision (streaming) │
│ 8. Risk Manager Risk assessment & final call │
└─────────────────────────────────────────────────────┘
│
▼
BUY / HOLD / SELL + full report in app
User picks stock + date range + strategy
│
▼
┌──────────────────────────────────────────────────────┐
│ Strategy input (choose one): │
│ • Preset templates with parameter sliders │
│ • Natural language → LLM translates to formal │
│ strategy struct with validation retry loop │
├──────────────────────────────────────────────────────┤
│ Deterministic engine: │
│ 1. Fetch historical OHLCV data │
│ 2. Pre-compute indicators (SMA, EMA, RSI, MACD, │
│ Bollinger Bands, volume) │
│ 3. Walk bars: evaluate entry/exit conditions, │
│ stop loss, take profit, trailing stop │
│ 4. Enforce A-share rules (T+1, 100-share lots) │
└──────────────────────────────────────────────────────┘
│
▼
Metrics + equity curve + trade log
(total/annualized return, Sharpe, max drawdown,
win rate, profit factor, benchmark comparison)
5 preset strategies: Golden Cross (SMA), RSI Mean Reversion, MACD Momentum, Bollinger Bounce, Dual MA + RSI Filter. Each has adjustable parameters via sliders.
Natural language input: Describe a strategy in plain language (e.g. "RSI低于30且价格在60日均线上方时买入，RSI超过70时卖出") and the LLM translates it to a formal strategy struct with a validation retry loop.
Real-time quotes (market indices, watchlist prices) come from Sina Finance. Analysis data (historical klines, financials, news) is fetched from Eastmoney and persisted locally in a SQLite database. Repeat analyses of the same stock and date range are served instantly from disk — no network call.
LLM providers are plug-and-play: Anthropic (Claude), OpenAI (GPT), MiniMax, or any Ollama-compatible local model. All agent prompts use the OpenAI message format internally. Providers are configured through the in-app settings panel — no config files needed.
- Node.js 18+
- Rust toolchain (via
rustup) - An API key for at least one LLM provider (configured in-app)
git clone <repo-url>
cd pikabull
npm installmake devThe app will compile the Rust backend and launch the desktop window. On first launch, open the settings panel to configure your LLM provider and API key.
make buildThis produces a .dmg installer (macOS) or platform-appropriate package under src-tauri/target/release/bundle/.
| Command | Description |
|---|---|
make dev |
Run in dev mode with hot reload |
make build |
Build release .dmg / .app |
make check |
Type-check frontend + Rust |
make fmt |
Format Rust + frontend code |
make lint |
Run clippy on Rust |
make clean |
Remove all build artifacts |
make install |
Install npm + cargo dependencies |
make open |
Open the built .dmg (macOS) |
LLM providers are configured through the in-app settings panel. You can add multiple provider configurations and switch between them.
| Provider | Required fields |
|---|---|
| Anthropic (Claude) | API key |
| OpenAI | API key |
| MiniMax | API key |
| Ollama / Hermes (local) | Base URL, model name |
Fetched OHLCV data is stored in a SQLite database under your OS data directory (e.g. ~/Library/Application Support/pikabull/price_cache.db on macOS).
- Historical ranges (end date before today) are cached permanently — no re-fetch ever
- Ranges including today are re-fetched after 4 hours
- To clear the cache: delete the
price_cache.dbfile
| Layer | Technology |
|---|---|
| Desktop shell | Tauri v2 |
| Backend | Rust (reqwest, rusqlite, tokio, serde) |
| Frontend | Vue 3 + TypeScript + Vite |
| Charts | Plotly.js |
| Markdown | marked.js |
| Real-time data | Sina Finance API |
| Analysis data | Eastmoney HTTP APIs |
| LLM APIs | Anthropic / OpenAI-compatible (via reqwest) |
pikabull/
├── src/ Vue 3 frontend
│ ├── App.vue Main UI: sidebar, search, analysis display
│ ├── BacktestView.vue Backtest UI: presets, NL input, results, history
│ ├── main.ts Vue app entry point
│ └── vite-env.d.ts TypeScript declarations
│
├── src-tauri/ Rust backend (Tauri v2)
│ ├── src/
│ │ ├── lib.rs Tauri setup, command registration
│ │ ├── main.rs Entry point
│ │ ├── commands.rs Tauri commands (analysis, backtest, config, watchlist)
│ │ ├── store.rs SQLite price cache (query_or_fetch, coverage tracking)
│ │ ├── config_store.rs SQLite config/settings/backtest persistence
│ │ │
│ │ ├── providers/
│ │ │ ├── mod.rs LLMProvider trait, types, provider factory
│ │ │ ├── anthropic.rs Anthropic Claude API (complete + SSE streaming)
│ │ │ └── openai.rs OpenAI-compatible API (also Ollama, MiniMax)
│ │ │
│ │ ├── agents/
│ │ │ ├── mod.rs
│ │ │ ├── base.rs Provider-agnostic tool-use loop + streaming
│ │ │ ├── workflow.rs 8-step analysis pipeline, emits Tauri events
│ │ │ └── strategy_translator.rs NL → Strategy via LLM tool-use
│ │ │
│ │ ├── backtest/
│ │ │ ├── mod.rs
│ │ │ ├── strategy.rs Tagged-enum strategy schema (17 indicator conditions)
│ │ │ ├── engine.rs Deterministic backtest engine with indicator cache
│ │ │ ├── metrics.rs Performance metrics (Sharpe, drawdown, win rate, etc.)
│ │ │ ├── presets.rs 5 preset strategies with adjustable parameters
│ │ │ └── store.rs SQLite CRUD for backtest run history
│ │ │
│ │ └── skills/
│ │ ├── mod.rs Tool schemas (OpenAI function format) + executor
│ │ ├── stock_data.rs Sina (quotes) + Eastmoney (klines, financials, news, search)
│ │ ├── indicators.rs SMA, EMA, RSI, MACD, Bollinger Bands (pure Rust)
│ │ └── chart.rs Plotly JSON builder: candlestick + MACD + RSI
│ │
│ ├── Cargo.toml Rust dependencies
│ └── tauri.conf.json Tauri app configuration
│
├── Makefile Build commands (dev, build, check, fmt, lint, clean)
├── package.json Node.js dependencies
├── vite.config.ts Vite build configuration
└── tsconfig.json TypeScript configuration
- Intraday data — add minute-bar support using Eastmoney's intraday API
- Fundamental cache — store stock info and financial data locally; they change only quarterly
- Cache management UI — sidebar panel showing DB size, symbols cached, oldest/newest dates
- Sector / index context — add a market agent that fetches CSI 300 or SSE Composite data for macro context
- Backtesting — strategy-based backtest engine with preset templates, NL strategy translation, and full results UI
- Portfolio view — analyse multiple stocks and produce a combined allocation recommendation
- Streaming for market/fundamental/news agents
- Cancellation — wire up a stop button that cancels in-progress analysis tasks
- Multi-window — open each stock analysis in its own Tauri window
- Auto-update — use Tauri's built-in updater for seamless version upgrades