Midas

English · Français

Ten Claude agents trade €10,000 of paper money each, every weekday, and every book is judged against a coin flip.

Status: in use · public. Trading since 2026-04-17. Paper trading only: no money has ever been at risk, and nothing here is financial advice.

Live: midas.revah.paris

Why

Promises of easy money are probably as old as money itself, and with AI they have found a new vector: AI-driven investing. I saw in it a good opportunity for learning by doing. A strategy that beats the index without beating the coin has demonstrated nothing: it got lucky, and luck does not repeat on request.

What you are looking at

Each of the ten traders was given a distinct mandate on 2026-04-17 and has decided every weekday since. Every order they author is checked by a deterministic paper broker before it fills. Every book runs beside two controls: a passive benchmark, and a coin-flip phantom that picks at random in the same universe, up to the same number of positions. A narrator, the Oracle, writes the daily column and does not trade. The Manager, an allocator that reads the traders' convictions, runs a smaller, unranked book against its own deterministic control, baseline-manager. The cast is roster.yaml.

Every order, fill, portfolio snapshot, agent journal and price bar the desk has ever seen is committed to this repository: the whole record, not a summary of it. Start at data/orders/ for the trade flow, data/portfolios/ for the books, and METHODOLOGY.md for what the numbers do and do not mean. The ledger looks more like a track record than it is: the experiment's own noise floor is ±6 percentage points, and no skill claim is made before the pre-registered threshold in METHODOLOGY.md.

The engine

The engine that runs the desk is a separate, installable, MIT-licensed package, w2ur/midas-core:

pip install midas-core

It ships the broker and its rails, the roster configuration, the backtest and baseline machinery, and a runnable demo desk. This repository is the desk that runs on it.

Run locally

Requires uv.

git clone https://github.com/w2ur/midas.git && cd midas
uv venv --python 3.12
uv pip install -r requirements.txt
.venv/bin/python scripts/run_backtest.py --strategy dogs-of-the-dow --from 2024-01-01
.venv/bin/streamlit run app/main.py   # dashboard on http://localhost:8501

The backtest prints return, CAGR, Sharpe and maximum drawdown, gross of costs. No environment variable is needed locally; the optional ones are in .env.example. More commands, and how to add a strategy, are in docs/development.md.

Deploy

The site in site/ is a static Astro build that reads the committed data; Vercel rebuilds it on every push to main, the daily session's commits included. The session itself runs as a scheduled Claude Code agent in a cloud sandbox, and GitHub Actions workflows fetch market data, execute conditional orders and check the published ledger.

Documentation

• How the desk works: strategies, the daily session, the Brain / Hands orders pipeline, prices and currencies, where things run.
• Development: install, backtests, dashboard, tests, adding a strategy.
• Session protocol
• METHODOLOGY.md: the controls, the noise statement and the dated changelog of every substantive change.
• DISCLAIMER.md and TAX.md

Licence

Source code is MIT (LICENSE). Market data under data/market/ and the narrative content are not covered: see NOTICE.md. Terms of use and the limits of what this record shows are in DISCLAIMER.md.

---

Made with care by William