convez.ai
SkillFinanceiroAvançado
orangejustin

OpenTrading

orangejustin/OpenTrading

Copiloto de trading macro-first, risk-first, sem keys, 100% local

Ver no GitHub
InicialPluginScore · de 0 a 100

Toolkit open-source e local-first para análise de trading de curto prazo integrado com Claude. Funciona sem API keys, priorizando macro economia e gestão de risco, com tudo rodando na sua máquina e sem dependências externas além da stdlib Python.

Instalação

$npx convez-ai orangejustin/OpenTrading

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Análise de setup de trading com decisões Long/Short/Wait baseadas em macro e gamma
  • Monitoramento de posições com dashboard web local e email pré-market automático
  • Aprendizado de política de risco e posicionamento de smart money sem expor dados

Veredito

Ferramenta sólida para traders que entendem análise macro e querem automation local sem dependências. Curva de aprendizado íngreme, mas arquitetura bem pensada (macro-first, risk-first) resolve problemas reais. Claramente educacional, não substitui pesquisa fundamentada.

README

A tradução para o português deste README ainda está na fila.

A local-first trading copilot — macro-first, risk-first, zero API keys.

📖 Documentation & tutorial →

License: MIT Python 3.9+ API keys: none stdlib only Claude Code: native

Preview · First principles · Features · Quickstart · Web dashboard · Daily email · Privacy

English | 简体中文

OpenTrading fuses macro, news, smart-money positioning, and options gamma into one opinionated read — then turns it into concrete action: a graded Long / Short / Wait board, sniper levels, and a CALL / PUT / NO-ACTION decision engine that follows a learned, risk-first policy. Everything runs on your machine: no SaaS, no keys, no data leaving your laptop. Pair the trading skill (the expertise) with small, dependency-free data CLIs (the live data), and let Claude drive both.

⚠️ Educational only — not financial advice. Trading carries substantial risk of loss.


Preview


First principles

Three rules the whole system enforces, in order:

#PrincipleWhat it means in the product
1Macro first, setup second, size third.Every read starts from rates/liquidity + dealer gamma + sentiment — a trade idea that fights the regime gets demoted, whatever the chart says.
2Risk before opportunity.No level without an invalidation: every plan carries a stop and a size derived from that stop; the event gate (FOMC/CPI/OPEX/earnings) can veto sizing entirely. NO-ACTION is a position.
3Local first, keyless first.Public no-auth endpoints, Python stdlib, everything on 127.0.0.1. The AI layer is optional garnish — three interchangeable engines, one of which is your existing Claude subscription with no key at all.

What you get

PieceCommandWhat it does
Market reportotFuses macro + news + smart money + options + your book → one regime read
Web dashboardot webLocal cockpit: tape · Macro & Flow · Strategy board · News + event gate · prediction desk per ticker · a Learn textbook · a 中文/EN toggle
Prediction deskot debate / ot rankBull vs bear on different engines → a committed verdict; one composite score ranks your whole book
Forecast conesot quant / ot forecastKeyless logistic P(up) + empirical cone; optional TimesFM foundation-model cone
Decision engineot decideCALL / PUT / NO-ACTION + conviction + zones, from the learned policy
Daily emailot email / ot schedulePosition-aware, Outlook-safe HTML pre-market brief via SMTP (EN or 中文)
Newsot newsFinancialJuice squawk (public RSS) — windowed, ticker-filtered, storable
Macro & flowot macro / ot smart / ot polyRates/liquidity bias · Fear&Greed · BTC funding · Polymarket crowd odds
Optionsot optionsPut/Call + dealer gamma (GEX) + gamma walls (CBOE)
Positioningot hl / ot whalesHyperliquid perp funding/OI · labeled-wallet ETH flows (keyless RPC)
Event gateot catalysts / ot earningsFOMC/CPI/PCE/NFP/OPEX + per-name earnings → size-up verdict
Quotesot quote / ot cn / ot cnpackNo-key quotes incl. pre/post/overnight + ^VIX; China A / HK · 涨停池
Data sanityot validate / ot privacy-auditCross-source quote check (Yahoo × CBOE) · pre-push secret/branch gate

Add --json to any tool for machine-readable output. Full help: ot help.


The prediction desk

The flagship: a personal short-term prediction desk — an ensemble of independent forecasters, fused by an LLM judge, scored by a learning loop. Deterministic SOP throughout (scripts build a frozen evidence pack → the LLM reads it; never agentic).

FORECASTERS  regression P(up)+cone · TimesFM cone · dealer gamma · macro · crowd odds · news
     │       (each an independent "analyst" — different math, different data)
     ▼
FUSION       ot debate  → bull vs bear on DIFFERENT engines → judge commits (5-tier verdict
             + entry + invalidation + time stop) · consensus strip (STAND-ASIDE on disagreement)
             · confluence ladder (a level named by 2+ methods is real structure) · ot rank
     ▼
LEARNING     ot reflect → every verdict journaled with its invalidation, graded on outcome,
             the lessons fed back into the next judge prompt

On every ticker page in ot web: dual forecast cones, the one-click ⚔️ debate, the confluence ladder drawn on the chart, and a ? on each module linking to the Learn tab — a cited textbook (with a full 简体中文 edition) on how each piece works and how to read it. The full computation reference is docs/METHODOLOGY.md.


Quickstart

60 seconds, no keys:

git clone https://github.com/orangejustin/OpenTrading
cd OpenTrading
bash install.sh        # puts `ot` on PATH + a health check — nothing to sign up for
ot                     # the morning read: macro + news + smart money + options + your book
ot web                 # the dashboard → http://127.0.0.1:8787

Don't want to touch your PATH? Skip install.sh and run in place: bin/ot …

New here? Start with the three hero workflows — morning read, is it safe to size up?, and grade my book — each is one command plus one prompt to Claude.

Ask Claude

Open the folder in Claude Code (or Claude Desktop) and just ask — the embedded short-term-trader skill activates automatically and pulls live data through ot:

  • "Give me my morning macro brief — calls or puts on QQQ today?"
  • "Any FinancialJuice news on NVDA in the last hour? Store it."
  • "NVDA broke $950 on volume, RSI 62, account $30k — how do I trade it?"

The skill enforces the house rules on every answer — the first principles above — and its workflows cover macro bias, news impact, trade setups, options, crypto sizing, the P&L journal, backtesting, and portfolio review.


AI engines — bring your own

The data panels are keyless forever. The optional AI layer runs on your choice of engine, switchable live from the dashboard header (or ot web --engine … --model …):

EngineKeyModels
GeminiGEMINI_API_KEY (free tier)gemini-2.5-flash / -pro
OpenRouterOPENROUTER_API_KEY — one key, any modelGLM 5.2 · DeepSeek v4 · GPT-5.5 · Claude · Grok · any slug
Claude Codenone — your existing subscriptionheadless claude -p (default / sonnet / opus / haiku; the exact resolved model is stamped on every run)
Codexnone — your ChatGPT/Codex subscriptionheadless codex exec in a read-only sandbox (appears when the codex CLI is on PATH)

Every analysis is stamped engine · model · latency · finish time (ET) and cached until you re-run. No engine configured? Everything still works — you just don't get the AI card.


ot web — the local dashboard

stdlib http.server + vanilla JS, no build step, bound to 127.0.0.1 (see the preview):

ot web                                          # http://127.0.0.1:8787
ot web --engine claude                          # boot on the no-key Claude Code engine
ot web --engine openrouter --model z-ai/glm-5.2 # boot on GLM 5.2
  • Dashboard — scrolling macro tape (TradingView deep links) · index cards · Macro & Flow (macro score · Fear&Greed · BTC funding · SPY dealer gamma + walls) · your watchlist.
  • Strategy — the action board: one deterministic ot decide card per name (Long/Short/Wait · grade · zones · stop), ~1s, no LLM.
  • News — 6h–7d window (deep windows merge the local archive), instant keyword filter, event-gate strip, and a 🧠 AI read of the tape (bias · drivers · portfolio tilt).
  • Ticker pages — instant keyless candlestick chart (with pre/post/overnight price) + the prediction desk (dual cones · ⚔️ debate verdict · confluence ladder) + per-name news, then ⚡ Analyze on demand; deep links like /#NVDA; optional TradingView chart embed.
  • Learn — a ? on every module opens the textbook: concepts, the models used, how to read each output, real annotated cases, and citations. Full 简体中文 edition.
  • 中文 / EN — the header toggle (or ?lang=zh) flips every label and the LLM output; point the dashboard at another roster with OT_WATCHLIST=path ot web.

Details: tools/web/README.md.


Daily pre-market email

A position-aware pre-market brief in your inbox every weekday — the same fusion as ot, written by Claude on your subscription and delivered as styled, Outlook-safe HTML: regime call, your book with levels, a graded Top-3 watchlist (Enter / Wait at an exact price), hedge plan, and the day's event gate.

cp .env.example .env       # set OT_SMTP_* + OT_EMAIL_TO  (Resend works with no 2FA)
ot email --dry-run         # confirm config (no send)
ot email                   # one-off send   ·   --lang zh for 简体中文
ot schedule email          # weekdays 08:30 local (macOS launchd) · `… email uninstall` to remove

macOS: launchd can't read repos under ~/Desktop, ~/Documents, or ~/Downloads (TCC) — keep the repo elsewhere (e.g. ~/OpenTrading). Details: tools/email/README.md.


ot decide — the policy in one call

ot decide TICKER --dte N turns the written policy into a single concrete call — CALL / PUT / NO-ACTION + conviction + zones — from no-key data:

ot decide QQQ  --dte 0     # 0DTE: fade-gap + VIX-confirm + skip-events + selectivity
ot decide NVDA --dte 5     # swing: momentum calls on names you read well

It encodes references/learned-strategy.md (selection > timing; a hard daily-loss stop; never size up after a loss). The Strategy board in ot web is this engine, fanned out across your whole book.


Privacy

Your holdings and secrets never enter git and are never part of any release:

WhatLives inStatus
Your positionswatchlist.jsongit-ignored — only watchlist.example.json is tracked
Email / API credentials.envgit-ignored — only .env.example is tracked
Fetched news, reports, briefsdata/git-ignored
cp watchlist.example.json watchlist.json   # then edit with YOUR positions
cp .env.example .env                        # then add your SMTP creds

The *.example files are placeholders; the real ones stay on your machine. The web dashboard binds 127.0.0.1 only. Never commit .env or watchlist.json.


Optional power modules

The core above is the plain tier: free, keyless, zero manual steps. These add more but are optional — nothing in the core depends on them.

  • TradingView (shipped) — bridge your TradingView Desktop app to Claude via the tradingview-mcp server, then ask "analyze MSTR with the TV data". The dashboard also has an opt-in TradingView chart embed per ticker. (ToS-gray; runs against your own logged-in client.)
  • IBKR (planned, tools/ibkr/) — Interactive Brokers via ib_async: quotes, option chains, and paper execution behind an explicit guard. Never auto-submits live orders.

Roadmap

Shipped history in docs/RELEASE_NOTES.md; detail in docs/ROADMAP.md. Recently shipped: the multi-engine debate, decision-log v2 (grade + lessons loop), ot whales, the confluence ladder / consensus fusion, ot rank, the Learn textbook, the 中文/EN toggle, and ot validate. On the horizon:

  • Calibration-weighted fusion. Once ~30 calls are graded, the judge weights each analyst by its track record instead of voting them equally.
  • IBKR paper execution. Debate verdicts auto-executed in a paper account behind an explicit guard — the on-ramp to (gated) tiny real sizing. Never auto-submits live orders.
  • Cross-device / cloud. An optional hosted tier (Cloudflare Pages + Workers + Supabase) so the dashboard is viewable from your phone — see docs/CLOUD.md.

Requirements

Python 3.9+ (standard library only; uses certifi if installed, else falls back to system curl). No keys, no paid feeds. ot auto-prefers uv when installed and otherwise runs on plain python3 — override with OT_PYTHON, disable uv with OT_NO_UV=1, inspect with ot doctor.


Credits & disclaimer

Built by @orangejustin. The multi-agent direction draws on TradingAgents.

Analysis is for educational purposes only — not financial advice. Markets are risky; size accordingly and do your own research.

Ver o original no GitHub

Bom para

Relacionados