StockAI

English | 简体中文

An AI equity research desk that runs on your own machine. StockAI gathers live news, prices, fundamentals and money flow for US stocks and Chinese A-shares, then puts AI on top of them: a fast sentiment read, a panel of 13 investor-master agents that each argue the case from their own philosophy, and a chat window that answers questions with citations back to the source. Built with Tauri 2 — one desktop binary, no server, and your API keys never leave the device.

Bring your own model. OpenAI, Anthropic, DeepSeek, GLM, or a fully local Ollama. This project hosts nothing and proxies nothing.

🌟 Features

🧠 AI analysis, in three depths

• Sentiment scan — collects recent news for a ticker and returns a bullish / bearish ratio with colour-coded factor cards. Deep Mode extracts full article bodies instead of just headlines.
• Deep Master Analysis — 13 investor-master agents (Buffett, Munger, Graham, Burry, Wood, Lynch, Fisher, Ackman, Pabrai, Taleb, Druckenmiller, Damodaran, Jhunjhunwala). Each is one LLM call carrying that investor's actual framework, and a synthesizer aggregates them into a consensus rating you can expand master by master. You pick who runs, and the panel shows the exact LLM-call count before you spend anything.
• Ask the filings — a chat panel grounded in the company's earnings reports, the news just scraped, and the quant metrics on screen. Every answer carries citations you can click back to the original.

Analysis never fires on its own. Switching tickers loads market data but spends no tokens — every LLM run is an explicit click. Models are assignable per role, so a cheap one can handle summarising while a strong one does the reasoning.

📊 Quant scoring, screening & backtesting

• Quant score (1–100) — a technical + fundamental composite, blended with a valuation read and pulled toward neutral by a volatility measure. When a dimension is unavailable it degrades to a weaker claim instead of inventing a number.
• Natural-language screener — type "main-board A-shares with ROE above 15% and P/E below 20", and it parses the conditions, sweeps the entire A-share market, then pulls fundamentals for the leading candidates. It shows you how it interpreted your sentence before spending time on it.
• Strategy backtest — run a strategy across the loaded history and get total return, annualised return, trade count and a buy-and-hold baseline, with entries, exits and the equity curve drawn on the chart.
• Master track record — every directional call a master makes is logged locally and marked to the current price, so hit rates and an equal-weight NAV curve accumulate forward from your very first analysis.

📈 Charts & market data

• Interactive K-line chart — MA / BOLL overlays, nine ranges (1D / 5D / 1M / 3M / 6M / YTD / 1Y / 5Y / All), a switchable sub-chart (MACD / RSI / KDJ / OBV / VWAP), log scale, forward/backward dividend adjustment, a comparison baseline, and the live price merged into the last candle during trading hours.
• Market overview — a persistent index bar (SSE / SZSE / ChiNext / S&P 500 / Nasdaq / Dow) that opens into industry and concept sector leaderboards (main-force net inflow, advancers/decliners, leading stock) plus the Dragon-Tiger list: the latest session's top net buys and sells with each listing reason.
• Company profile (F10, A-shares) — corporate overview, latest-period revenue breakdown by product / industry / region with gross margins, holder-account count, top-10 tradable shareholders and sector tags. Fetched only when you expand it.

💼 Your own book

• Portfolio tracking — real positions (shares + weighted-average cost) with live market value, unrealised P&L, day P&L and position weight; adding to a holding merges the cost basis by share-weighted average. Currencies are totalled separately — there is no FX data source here, and summing CNY with USD would produce a meaningless number. Positions without a quote are excluded from totals and named explicitly, never filled in with zeros.
• Watchlist & price alerts — quick-add from the sidebar, with optional upper / lower price alerts delivered as system notifications.
• Analysis history — every analysis, deep analysis and screen is archived to a local SQLite database and reopens in full.

⚙️ Built to keep working

• Redundant data sources — quotes, K-lines, fundamentals, sector boards, fund flow and market snapshots each fall back across Tencent / Eastmoney / Sina / Yahoo, so a single upstream outage doesn't blank the app. News keeps working where Google and Yahoo are unreachable, through an Eastmoney source that depends on neither domain.
• Local-first, key-safe — settings and API keys live in the OS-level store. Keys reach the analysis process through a 0600 temp file, never through command-line arguments where any process listing would expose them.
• Three languages — 简体中文 / English / 日本語, switchable at runtime and carried all the way down into the LLM prompts.
• Signed in-app updates — v0.15.0 and later update themselves.

🏗️ Architecture

Three layers with strictly unidirectional dependencies: UI → Tauri Core (Rust) → Sidecar (Bun).

1. Frontend (src/) — React 19 + TypeScript + Vite. All IPC goes through one entry point; local history is SQLite via tauri-plugin-sql.
2. Tauri Core (src-tauri/) — Rust. Exposes exactly one command, invoke_sidecar, which knows no business logic: it swaps sentinels for temp-file paths (config, oversized payloads) and cleans them up. Adding a capability therefore needs no Rust change.
3. Sidecar (sidecar/) — a compiled Bun binary, spawned per call and fully stateless. Scrapes news (plain-fetch strategies first, Playwright only when they come up empty), runs the quant computation, and talks to the AI providers.

shared/ is the single source of truth across all three: DTO types, the CLI action manifest, provider profiles and market detection.

Full contract: .claude/rules/architecture.md.

📦 Installation

Pre-built binaries are on the Releases page.

Then open the app → Settings (top right) → enter the API key for your provider of choice → Save.

macOS — "StockAI is damaged" error

macOS Gatekeeper blocks apps that aren't notarized by an Apple Developer certificate. Run this in Terminal to clear the quarantine flag:

xattr -cr /Applications/StockAI.app

Then open the app normally. This is safe — the app contains no network backdoors and the full source is auditable in this repository.

Why this happens: apps downloaded from the internet receive a quarantine attribute. Without an Apple code-signing certificate, macOS reports "damaged" instead of the usual "unknown developer" prompt.

Windows — SmartScreen warning

Click More info → Run anyway. This appears for any unsigned executable.

Linux (.deb)

sudo dpkg -i StockAI_*_amd64.deb

Requires WebKitGTK (pre-installed on most GNOME-based distros).

---

🚀 Building from source

Prerequisites

• Bun — package manager and Sidecar runtime. Install Bun
• Rust — for the Tauri core. Install Rust

1. Install dependencies

bun install

2. Start the dev environment

bun tauri dev

This compiles the Sidecar binary for your machine first, so there is no separate build step.

3. Package a release build

bun tauri build

To cross-compile the Sidecar for another platform explicitly:

BUN_TARGET=bun-windows-x64 bun sidecar/build-script.ts

Valid targets: bun-darwin-arm64 · bun-darwin-x64 · bun-linux-x64 · bun-windows-x64. The output lands in src-tauri/bin/ named with the Rust target triple Tauri expects.

🧪 Testing

bun run test              # full offline suite — frontend Vitest + Sidecar Bun tests
bun run test:integration  # the above, plus live-network assertions against every data source
bun run test:e2e          # boots Vite in a real browser and asserts the chart actually renders
bun run format            # Biome formatter
cd src-tauri && cargo test # Rust core

The default suite is entirely offline — scraping and parsing are decoupled so every network call has an injection point. Because offline fixtures never change, they are structurally blind to an upstream redesign; each external data source therefore also carries a shape assertion in a *.integration.ts file, run daily by CI.

🛠️ Tech Stack

• Desktop: Tauri 2 (Rust), SQLite via tauri-plugin-sql, signed auto-updates
• Frontend: React 19, TailwindCSS 4, Lightweight Charts v5, Lucide Icons
• Sidecar: Bun, Playwright, node-html-markdown
• AI: OpenAI SDK, Anthropic SDK, Ollama SDK (DeepSeek / GLM over the OpenAI-compatible protocol)
• Tooling: Biome, Vitest, Lefthook

📅 Development Conventions

• Code comments are written in Chinese (project preference).
• UI component files stay under 200 lines — complex logic is extracted into hooks.
• Unidirectional dependencies (UI → Core → Sidecar), with shared/ as the only cross-layer source of truth.
• Parsing logic lives apart from the network layer and must be covered by offline unit tests.

🤝 Contributing

Contributions are welcome! Please read the Contributing Guide to get started, and our Code of Conduct. Found a bug or have an idea? Open an issue or start a discussion.

⚠️ Disclaimer

StockAI is a research and educational tool. Its output — including AI-generated ratings, master consensus, quant scores and backtest results — is not investment advice. Market data comes from public third-party sources and may be delayed, incomplete or wrong. You alone are responsible for your investment decisions.

📄 License

MIT © hyhmrright