My most advanced AI project
Claude NY.
An AI agent that runs a trading research lab around the clock: it turns ideas into MetaTrader 5 strategies, tests every one on real tick data, deploys only what passes, and guards a live portfolio with real money on it.
Showcase, not for sale
What it is
Claude NY is my autonomous trading research and operations system. A Claude agent (Claude Opus, by Anthropic) lives on a Windows server in New York and never clocks off. I set the rules and the goals in plain language. It does the engineering: it reads trading ideas from free courses, YouTube strategies, published papers and my own older Expert Advisors, writes each one as an MQL5 Expert Advisor, and sends it to a second machine in Perth that tests it on real tick data.
Only strategies that clear a strict bar go live. From then on the agent watches them, protects them and keeps improving them, writes to the broker on my behalf, and reports to me. I can reach it from my phone at any hour through Claude Code Remote Control.
Architecture
Two machines with different jobs, a git repository as the nervous system, and AWS for data and mail.
The brain and the live desk
- Live MetaTrader 5 terminal, one chart per strategy (99, the platform's maximum)
- Python risk daemon guarding every open trade
- Watchers streaming events into the agent
- The Claude Code orchestrator session
The muscle
An Acer Predator Helios 300 bought in Canada in 2020 and upgraded by hand: Intel Core i7-10750H (6 cores, 12 threads), 64 GB DDR4, RTX 2070 Max-Q, and two 4 TB NVMe drives plus a 1.92 TB SSD, almost 10 TB in all.
- 9 MetaTrader 5 Strategy Tester copies in parallel
- A standing job queue that never runs dry
- Every test, study and machine learning experiment runs here
- New York only plans, trades live and analyses
- Plus a read-only data terminal feeding the market-data store
Data and voice
- S3 market-data store: MT5, Binance and Dukascopy history as Parquet, one file per month, updated incrementally
- Secrets Manager for the mail sender, and HELLO's mailer for reports and alerts by email
The whole loop in one line
The agents
One orchestrator holds the whole picture and makes deploy and retire decisions inside my standing rules. Background agents, each with exactly one job, do the rest. They talk through messages and shared files; a deploy lock and a restart gate make sure only one of them touches the live terminal at a time.
How a strategy earns real money
- Idea. A course lesson, a YouTube strategy, a paper, or one of my old EAs.
- Code. Claude writes it as an MQL5 Expert Advisor, with the shared safety modules compiled in.
- Queue. It goes to Perth as a git commit and joins the standing queue.
- Test on real ticks. One week, one month, one year, then an out-of-sample year. Crypto is split into weekdays and weekends; everything is split by weekly open, weekly close, the rest of the week and news.
- The bar. All four must hold, or it does not go live:
- Deploy. The auto-deployer puts it on a live chart. If the profile is full, the weakest chart is retired first.
- Grow carefully. A lot-size sweep: a bigger lot only if every smaller size has passed too.
Risk controls
Monitoring and recovery
Watchers stream every trade close, risk-daemon event, Perth failure, idle alarm and lot-sweep verdict straight into the agent's session. The server reboots daily on purpose: a resume checklist and a logon hook bring every watcher and agent back. Scheduled guards restart the risk daemon if it ever dies, a hygiene task compacts git storage every 4 hours, and if Claude's login is ever revoked I get an email.
Every rule I set and every lesson it learns is written to persistent memory files, so its behaviour survives restarts instead of starting from zero each morning.
Lessons that became rules
Every incident changed the system, not just the moment.
Giving back
Building Claude NY meant reading other people's MetaTrader tools closely. Where one could hurt a live account, I wrote it up for the maintainer, with the fix.
- Merged, 5 Oct 2026. claude-trading-skills #472 (a Claude Code skills collection with almost 3,000 stars): its backtester could silently pick up the live trading terminal and close it. Fixed in #473: a terminal path must now be given explicitly.
- Proposed. metatrader-mcp-server #52: a "seatbelt mode" for live accounts, with a read-only switch and a confirm step.
- Proposed. mcp-metatrader5-server #99: read credentials from the environment so a password never passes through an AI chat.
- Proposed. mt5-trade-split-manager #13: refuse to start without an API key.
Components
- Live MT5 terminal: 99 charts of custom EAs (HourOne, the Learn course engine, CryptoTrend, WeekendRevert, MondayOpen, MachineGun and more) across forex, gold and silver, indices (NAS100, US500, US30, GER40, UK100, JPN225), stocks and crypto (BTC, ETH, SOL, LINK).
- first_lock_daemon (Python): virtual profit locks, loss-stop removal, news, weekend and open guards, swap-aware take-profit shift, request throttle.
- KT_LiveExit (MQL5): the same exit logic inside the Strategy Tester, so tests match live behaviour.
- KT_StopGuard, KT_BarGuard, negative-trades guard: shared safety modules compiled into every EA.
- Perth standing queue: 9 tester workers, staged runs, self-healing git sync, and a learning lane that keeps course tests at the top.
- Auto-deployers: apply passing builds and lot sizes to live charts by magic number, with backups and a lock.
- Watchers: live deals, risk log, Perth results, lot sweep, limit entries.
- Market-data store on AWS S3, and the HELLO mailer for reports.
Best week so far
In its best week, 26 Sep to 3 Oct 2026, on a 14,000 AUD deposit made on 23 Sep 2026, Claude NY made +447.81 AUD net on closed trades: about +3.2% of the deposit in seven days.
Best week, not an average: closed trades only (profit plus swap plus commission), open positions not included, and the account was eleven days old when this was written (5 Oct 2026). Past results say nothing certain about future ones.
Milestones
2025 My first live algorithmic trading system with Claude models in the loop, and hard risk limits.
Mid 2026 The first Claude-built EAs and studies on the New York server.
Sep 2026 The Perth testing machine linked; the learning programme starts: every free course and YouTube strategy tested as an EA.
30 Sep 2026 The never-close-negative rule and virtual profit locks.
2 to 3 Oct 2026 Re-entry and predicted-entry studies live; tester and live exits unified; the profile reaches 99 charts.
4 Oct 2026 Request throttle and restart gate; the AWS market-data store started; batch 22 (Heikin Ashi, Hull MA, Ehlers Fisher) live.
Why it matters to me
Thirty years of running technology taught me that the hard part is never the clever bit. It is the guard rails, the recovery, the rule that survives the next incident. Claude NY is that lesson with an AI doing the engineering: I decide what must never happen, and the agent builds, tests and runs everything inside those lines. It is a showcase of how I work, not a product and not advice.
Trading carries real risk of loss. Nothing on this page is financial advice.