Claude-native components live under .claude/ — never at the repo root. Slash commands go in .claude/commands/, skills in .claude/skills/<name>/SKILL.md, subagents in .claude/agents/. A root-level ./commands/, ./skills/, ./agents/, ./slash-commands/, or ./subagents/ is a bug — Claude Code will not discover them. Never add .claude/ to .gitignore; always commit it (use .gitkeep if empty).

This repository is a template provided as part of the New Repo From Template Plugin for Claude Code.

Claude FX Pair Analyst Template

A Claude Code workspace for analysing foreign exchange pairs — historical data retrieval, quant/econometric analysis, volatility modelling, support/resistance identification, grounded geopolitical framing, and PDF report generation.

What this template gives you

• A canonical tool manifest (toolkit.json) — curated FX-relevant Python libraries, APIs, and MCPs. Skills and commands reference tools by id.
• A grounding layer (research/fxpairs/) — durable per-pair profiles (drivers, liquidity pockets, vol character, hypothesis seeds). Built once, refreshed over time, read by every subsequent analysis.
• Workflow commands that chain the toolkit:
  • /profile-pair <PAIR> — durable reference profile
  • /analyze-pair <PAIR> [timeframe] — full analysis + PDF
  • /event-study <PAIR> <DATE> <LABEL> — dated event reaction
  • /compare-pairs <PAIR1> <PAIR2> … — multi-pair comparison
  • /cointegration-scan [BASKET] — stat-arb candidate scan
  • /vol-regime <PAIR> — volatility regime characterisation
  • /test-hypothesis <PAIR> "<claim>" — empirical hypothesis testing
• A shared Typst reporter (.claude/skills/generate-fx-report/) — consistent PDF output across every command.

Who it's for

Analysts, traders, and researchers who want a Claude-driven FX desk — opinionated about epistemics (measured vs presumed, falsifiable signals, cited claims) and pragmatic about tools (install-on-demand, lean dependencies, no crypto-only tooling).

Not for

• Retail auto-trading signal generation
• Crypto-only analysis (see goat-sdk or ccxt-focused templates)
• High-frequency execution (see nautilus_trader)

Quick start

1. Clone or scaffold from this template.
2. Open in Claude Code from the repo root.
3. For a new pair: /profile-pair EURUSD (builds grounding).
4. Then: /analyze-pair EURUSD 5y (produces PDF in reports/).

Requirements

• Python 3.10+
• Typst CLI (cargo install --locked typst-cli) for PDF rendering
• Optional API keys for enhanced workflows:
  • FRED_API_KEY — macro series (recommended)
  • ALPHAVANTAGE_API_KEY — intraday FX
  • OANDA_API_TOKEN + OANDA_ACCOUNT_ID — broker-grade data
  • NIXTLA_API_KEY — TimeGPT forecasting
  • PERPLEXITY_API_KEY — geopolitical synthesis

The toolkit installs tools on demand — you won't pay for anything you don't use.

How it's organised

toolkit.json                   # canonical tool manifest (libs, APIs, MCPs, patterns)
CLAUDE.md                      # the agent's role and working rules
research/fxpairs/              # durable pair profiles — grounding layer
data/                          # cached raw data
charts/                        # generated charts
reports/                       # final PDFs
context/from-human/            # your specific instructions
context/from-agent/            # the agent's working notes (task_plan / findings / progress)
.claude/commands/              # slash-command entry points
.claude/skills/                # shared skills (e.g. generate-fx-report)

Epistemic discipline

This template enforces a few rules that distinguish it from generic "AI trading bot" scaffolds:

• Every claim cites its source. Historical and narrative claims need a URL or a dataset. Unsourced claims are marked [unsourced — analyst inference].
• Measured and presumed are separated. Vol numbers are measured from data. "EURUSD is driven by rate differentials" is presumed unless tested — /test-hypothesis exists for that.
• Signals are falsifiable. Every logged signal carries strengthen / weaken / falsify conditions.
• No p-hacking. /test-hypothesis requires the test to be specified before looking at the result.

License

MIT. Use it, fork it, improve it.