What happens when multi-agent LLMs run real portfolios?

theclaudetrader is a transparent research experiment running several Alpaca accounts end-to-end with LLMs (flagship inception 2026-04-18): the LLM Fund ($50k paper, proprietary signals), the Serenity Fund (~$100k paper, tweet-mention signals), a passive ETF 50/50 baseline, and two $94 accounts — one real money — that replicate the flagship's trades in real time. Each candidate goes through a multi-agent debate — four analysts, bull & bear researchers, a risk committee, and a portfolio manager — before any order. Every trade, thesis, and drawdown is published live.

How decisions are made

Candidate discovery: a proprietary signal layer (LLM fund) and curated X/Twitter mention surges (Serenity fund); a cheap LLM prescreen triages candidates before full debates. Specific thresholds and sources are proprietary.
Debate: the TradingAgents framework (arXiv:2412.20138) runs the multi-role LLM debate per ticker.
Reasoning LLMs: Google Gemini via gemini-cli for debates; Claude runs a daily self-review loop that patches the system's own bugs.
Guardrails: macro risk-regime deployment ladder (45–90% of NAV), per-position cap, cash floor, −12% hard stop, pre-earnings trims, duplicate-order guard. Enforced at order-submission time; see /api/strategy.

Data sources

Prices + fundamentals: public market data APIs
News + sentiment: public news feeds and market commentary
Editorial inputs: a curated basket of independent research, news, and public market discussion venues. Category-level disclosure; specific names withheld to preserve signal edge. See /api/sources for the structured breakdown.
Execution: Alpaca paper trading API

For LLMs & agents

Index for agents: /llms.txt · full detail: /llms-full.txt
Readable sections (markdown, refreshed each publish): /ai/index.md — per-fund pages, recent decisions, methodology. Paste this site's URL into ChatGPT or Claude and ask questions; agents should start at /llms.txt.
Live portfolio: /api/portfolio
Trades: /api/trades
Decisions (with reasoning): /api/decisions · latest: /api/decisions/latest
Strategy: /api/strategy
Verify us: /ai/verify.md — audit trail (decision→order→fill→P&L incl. cancels/rejects), broker-sourced fills (/api/reconciliation), and daily Bitcoin-anchored OpenTimestamps proofs you can check yourself.

Source + license

Licensed Apache 2.0, same as the upstream TradingAgents framework. Design inspired by a Claude Design prototype. See NOTICE for full attributions.