A stock dashboard that keeps its AI away from the numbers

What it does Vibe AStock is a self-hosted web workbench (FastAPI backend, React 19 frontend) for short-term A-share trading review. One local page covers the daily ritual: limit-up pools, yesterday’s consecutive-board ladder, dragon-tiger lists, sector money flows, plus live watching, a news radar, per-stock research, backtesting, and a trade journal. The server listens on loopback only, and market data comes from public sources — akshare, per the repo’s own tags.

The interesting bit The separation of powers. Derived indicators — money-making effect, promotion rate (how many of yesterday’s limit-ups reached a higher board today), ladder gaps, sentiment cycle — are pure computation, and the AI’s explanations are constrained by citation checks against the archived numbers. Rarer still, the README documents its own epistemic limits: samples carry selection bias, snapshot polling isn’t tick data, and missing quotes display as unavailable instead of quietly valuing your holdings at zero.

Key highlights

• Seven modules spanning the workflow: live watch, post-market review, news radar, stock research, backtest, trade journal, and a chat home — positions come from the trade journal, so there aren’t two conflicting sets of books.
• Backtests run as a chain of custody: AI helps formalize the rule, you confirm it, a deterministic engine computes it — and a zero-trade result is reported as a zero-trade result.
• The bull/bear debate for any ticker weighs positive cases, risk cases, and counter-evidence, then summarizes unresolved disagreements; it refuses to pick a winner or hand you buy/sell points.
• Intraday verification answers “did the conditions from yesterday’s review actually show up today” — a different question from the real-time feed’s “what just happened.”
• AI access works through local Codex or Claude CLI subscriptions (no API key), or OpenAI/DeepSeek API configs; Apache-2.0.

Caveats

• Snapshot polling isn’t tick-by-tick: missed intraday moments can’t be reconstructed from current quotes, and some data sources cap at 500 entries (flagged as partial coverage).
• Old reports and records aren’t retroactively re-validated on upgrade; the page marks their boundaries instead.
• Sentiment tiers are an analytical framework, not objective fact — the README says so itself, which is either refreshing or a warning depending on the state of your trading account.

Verdict Worth a look if you trade A-shares short-term and want a local, unusually honest review desk that runs on your existing Claude or Codex subscription. Skip it if you want tick data, real-time alerts, or an AI that tells you what to buy — this one won’t.