superwrldcoin/trading-agent

An **analysis-only** trading research agent built on Claude Code. It reads 4H market structure off
previous day/week/month liquidity levels (PDH/PDL, PWH/PWL, PMthH/PMthL), grades conviction with
**EMA + VWAP**, and checks position math (R:R, liquidation, sizing), all from live public data.
It never places trades and never asks for or stores API keys.
Watchlist (from memory/universe.yaml): BTC/USDT, BCH/USDT (OKX, yfinance backup), XAU/USD (OKX Tether Gold proxy,
cross-checked against spot), SI (COMEX silver futures), SLV, GLD, MSFT (yfinance), plus 5 macro drivers and 8 pair instruments.
> Not financial advice. Outputs are analysis; you make and execute every decision.
Ask it something
| Interface | How | What runs |
|---|---|---|
| Web page | double-click start-ui.cmd (or python app/server.py), then open http://127.0.0.1:8765 | **Quick check** (no AI, ~30s on first fetch) or **Full agent** (Claude Code with every skill and memory file, a few minutes). Live step progress, cancel, run history, price ladder, settings drawer (theme, animations, defaults, shortcuts) |
| Terminal | .\ask.ps1 "BTC long 20x entry 98,400" / ./ask.sh ..., add -Agent / --agent for the full agent | same |
| Claude Code | cd trading-agent && claude, then type the question | the agent itself (interactive) |
The **quick check** parses the request, fetches 15M/1H/4H/1D data, and returns:
- a verdict
- EMA + VWAP conviction for the side
- structure
- a stop and targets derived from structure (or yours)
- a milestone table with R, P&L/ROE, and liquidation
- flags, such as liquidation hit before the stop, an entry far from the market, or R:R under 2
The **full agent** runs the quick check, then applies the skills, adds the event check and a
P(T1 before stop) judgment, and logs the session.
The full agent needs the Claude Code CLI logged in on the machine.
Run claude once inside this folder and accept the trust prompt. No API key is stored.
Everyday commands
After a trade closes, in claude: post-mortem: BTC long, entry ..., exit ..., size ..., stop ..., 20x, opened ..., closed ..., thesis was "...".
Review the proposed MEMORY_UPDATE, approve it, then commit memory/journal.md. Trades, sessions, and positions stay private (gitignored).
Setup
Layout
| Path | What |
|---|---|
| CLAUDE.md, AGENT.md | Agent rules: role, write scope, tools and fallbacks, session workflow, report format |
| skills/ | Procedures with worked examples: report-format, levels-and-entries, multi-timeframe-momentum (EMA + VWAP conviction), market-structure, risk-and-sizing, leveraged-position-math, leverage-volatility-check, rules-check, trade-plan, portfolio-exposure, stress-test, options-greeks, bear-case-review, data-ingest, macro-and-catalysts, post-mortem |