tradingagents

TradingAgents — Substanz Edition (v2.1)

A Claude Code plugin that runs a substance-maximizing, multi-perspective stock research workflow on any ticker. 4 base analysts run in parallel, then 13 famous-investor personas (Buffett, Burry, Wood, Druckenmiller, Taleb, …) weigh in, a bull/bear pair debates the case, a trader writes a quant-backed plan, and a risk manager signs off using ATR + Kelly hard caps.

Three workflow modes (/quick, /analyze, /deep) × five investor styles (value / growth / macro / contrarian / council). Includes a backtester, a continuous watch loop, a multi-ticker compare, and a read-only Next.js dashboard.

⚠️ Research, not financial advice. Output is for educational use only.

---

Credit & lineage

This project is a port + extension of two upstream projects:

• TauricResearch/TradingAgents — original LangGraph framework (agent role taxonomy, workflow shape, output conventions). The Claude-Code-native orchestration and MCP server are original to this repo.
• virattt/ai-hedge-fund — inspired the 13 famous-investor persona pattern and the JSON Signal Footer schema for structured agent outputs.

Both upstream projects and this plugin are released under Apache License 2.0. See LICENSE and NOTICE for full attribution.

---

What it does

        ┌── fundamentals-analyst (sonnet) ──┐
parallel├── technical-analyst    (sonnet) ──┤
        ├── news-analyst         (sonnet) ──┤
        └── sentiment-analyst    (sonnet) ──┘
                       ↓
        13 persona subagents in parallel (sonnet)
        Buffett · Graham · Munger · Pabrai · Burry · Taleb · Ackman ·
        Wood · Lynch · Fisher · Druckenmiller · Damodaran · Jhunjhunwala
                       ↓
        bull-researcher  ⇄  bear-researcher        (opus, mailbox debate, max 2 rounds)
                       ↓
        trader (opus) ─── reads ALL JSON signals ─── writes plan ───►
                       ↓
        risk-manager (opus) — uses atr_stop + kelly_position_size hard caps
                       ↓
                Final BUY/HOLD/SELL + position sizing + entry/stop/target

Every agent appends a JSON Signal Footer to its Markdown report; the trader does confidence-weighted aggregation across all signals before writing the plan.

---

Modes

Command	Analysts	Personas	Bull/Bear	Risk-Loop	Cost	ETA
/quick TICKER	2	3	—	—	$1–2	2–3 min
/analyze TICKER	4	5	1 round	1 iter	$3–5	6–8 min
/deep TICKER	4	13 + DCF	2 rounds	2 iter	$7–15	10–15 min

Styles (which personas run)

Style	Personas
value	buffett, graham, munger, pabrai, burry
growth	wood, lynch, fisher
macro	druckenmiller, damodaran, jhunjhunwala
contrarian	burry, taleb, ackman
council (default)	all 13 in deep, top 5 in standard, top 3 in quick

Pass with --style: /analyze NVDA --style value, /deep TSLA --style contrarian.

Other commands

• /backtest TICKER --from 2024-01-01 --to 2024-06-30 [--style value] — historical lite-workflow iteration, returns Sharpe / Sortino / Max-DD / win-rate vs buy-and-hold baseline.
• /watch TICKER --interval 1h [--alert-on signal-flip|price-break|news] — continuous monitor via ScheduleWakeup; max 24 iterations by default.
• /compare T1,T2,T3 [--mode quick] — parallel multi-ticker analysis with side-by-side persona heatmap and ranked top picks.

---

Install

1. Add the marketplace and install

/plugin marketplace add namikos/claude-tradingagents
/plugin install tradingagents@namikos-tradingagents

Claude Code will prompt for your Alpha Vantage API key (stored in your OS keychain via the plugin's userConfig). Free key in 30 sec at alphavantage.co/support/#api-key.

Optionally also set a Financial Datasets API key for premium fundamentals (skipped if absent).

Optional v2.1 API keys — all free-tier, all opt-in (tools return a clean error if a key is missing instead of crashing):

Key	Source	Powers
fred_api_token	FRED API (free)	fred() macro time series
fmp_api_token	Financial Modeling Prep (250/day)	earnings_transcript()
finnhub_api_token	Finnhub (60/min)	analyst recs + earnings/IPO calendar
reddit_client_id + reddit_client_secret + reddit_user_agent	Reddit OAuth app	reddit_mentions() PRAW search

2. Enable Agent Teams (one-time)

Add to your ~/.claude/settings.json:

{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}