simonlin1212/Vibe-Research
FinanceA local research workbench for China A-share, US and HK stocks, combining market data, AI-assisted company research, backtesting and a personal research library.
What it is
Vibe Research is a locally run investment research workbench covering China A-shares, US stocks and HK stocks. It puts index and quote monitoring, watchlists and a transaction-ledger portfolio, daily market review, industry news and signals, six-stage single-stock research, long/short debate, backtesting, a report library and a research log into one browser UI. Research records and imported documents stay on the user's own machine, and AI work is done through a subscription CLI already logged in locally (Claude Code, Codex, WorkBuddy/CodeBuddy) or any Chat Completions-compatible model API supported AI.
The data layer is a registry of 168 endpoints across 35 layers, spanning quotes, K-lines, financials, consensus estimates, announcements, funds flow, southbound capital, AH premium, options, indices, futures and commodities, event-driven filings and macro data endpoint catalog. A-share and global data are synced from the sibling projects a-stock-data and global-stock-data.
First run
Requirements: Windows 11, macOS or Linux; Node.js ≥ 22.18 (with TypeScript support enabled, recommended 24 LTS); Python ≥ 3.11 (3.12 verified); Git. The setup script does not install any AI CLI.
From the repository root, on macOS/Linux:
git clone https://github.com/simonlin1212/Vibe-Research.git vibe-research-agent
cd vibe-research-agent
scripts/setup
scripts/start
On Windows use scripts\setup-windows.cmd and scripts\start.cmd. Setup creates .venv, installs Node and Python dependencies and initializes the private data directory. Start launches both the local API and the UI and opens http://127.0.0.1:5930; if 5930 is occupied, the actual port is printed. Then open "接入 AI" (Connect AI) and pick a locally logged-in subscription CLI or enter an API base URL, model name and key, and run "测试并保存" (test and save). Browsing works without AI; research, debate and backtests need it.
There are also portable zip builds for Windows x64 and macOS arm64/x64, with bundled Node and Python and .sha256 checksums, on the Releases page.
For MCP use inside Claude Code, Codex or Cursor, register orchestrator/src/mcp.ts with the tool's own mcp add command; 7 of the 8 exposed tools are read-only, and starting a research run always uses the backend default provider, not the UI selection.
Constraints
- Research stages run without network or shell access; data is fetched by controlled scripts and derived numbers go through a deterministic calculation library. Missing data is recorded as a gap rather than zero.
- Subscription integrations use your own local CLI login; API keys entered in the browser are stored in
localStoragein plain form, so the tool should be used on a trusted personal machine. - HK/US six-stage research has narrower verification than A-shares: one full run each with a Claude Code subscription was done, and HK still had three uncited numbers in the test report.
- Optimizing for LAN or cloud access is optional and experimental; LAN mode shares the single-user backend without account isolation and uses plain HTTP, so it must stay on trusted networks. Remote access is macOS/Linux only.
- The portable packages are unsigned, so Windows SmartScreen and macOS Gatekeeper will block first launch; the macOS workaround is removing the quarantine attribute.
- Output is limited to data, frameworks, scenario probabilities and decision checkpoints; no position sizing, price targets or stop levels. Third-party public data may be delayed, missing or wrong.
- MIT licensed.
HotRepo's take
The project's center of gravity is its evidence discipline rather than the chat interface: each research run persists report.md, evidence.json, calculations.json, conflicts.json and a manifest, and numbers in reports that cannot be traced back are replaced with a placeholder. That design, plus the explicit market-by-market caveats for HK/US financials and the honest labeling of untested model API templates, is more than most stock-research side projects offer. Its weakest documented points are equally concrete: no real-machine Windows acceptance, unsigned portable builds, and HK/US research validated on only one run per market. On HotRepo it was recorded on the 2026-11-16 TypeScript daily list at 23 stars gained that day and position 10, with 3,738 stars and 704 forks overall; that single snapshot shows it drew attention in the TypeScript ranking but does not by itself establish sustained growth.
Written by AI from the project's README, updated . The project's own docs are the source of truth.
You might also like
- HKUDS/Vibe-TradingA Python agent for market research, backtesting, paper trading and opt-in broker execution.
- Fincept-Corporation/FinceptTerminalA native desktop finance terminal for retail and academic analysts who want Bloomberg-style research without a Bloomberg seat.
- ZhuLinsen/daily_stock_analysisAn LLM-driven stock analysis system that scores watchlists across six markets and pushes daily decision dashboards.
- shy3130/tick-stock-panelSelf-hosted A-share quant workbench for custom strategies, backtesting, market monitoring and LLM-assisted analysis.
- ErTasselli/OpenTerminalRun a local, keyboard-driven market terminal for US stocks, crypto, macro data and a small portfolio tracker.