Hi everyone!
I’ve been building a sort of Alpaca dashboard for about three months and it’s at a point where I’d rather have people poke at it than keep polishing in private.
PixelTrader is a desktop app that turns the Alpaca API into a 1980s-style trading floor. Eight brokers each cover a GICS sector. Every round, each one reads real price data, picks something it wants to buy or sell, and gives a one-sentence argument for it. You approve or you pass. They hold positions through drawdowns, take profits early, get things wrong, and carry that into the next round. A broker who loses enough gets fired and replaced.
The reasoning, that at times, can be semi flawed is the point. It allows for the player to act as a real manager, directing the brokers and the floor as a whole in the strategies the player feels are sound through the “coach” button.
How it’s wired up:
- Market data and execution both through Alpaca, with the user’s own keys. Keys live in Windows Credential Manager and never leave the machine, there’s no server, so I couldn’t see them if I wanted to.
- Three modes: sim, paper, and real. Sim replays a real 90-day historical window rather than reading live prices, so it runs at any hour and doesn’t need the market open. It’s the same engine and the same limits, just not reaching a brokerage.
- Every order requires explicit human approval in real mode. No autonomous mode, no overnight setting, no toggle to turn the approval step off, it’s structural rather than a preference. Real trading also requires arming the floor from the banner, and arming resets on every close.
- Hard caps that can’t be disabled: ~13% of equity per position, six positions max, a 15% cash reserve floor, a 30% day-loss stop per broker, and a 50% account breaker that survives a restart until acknowledged.
A few things I learned that might be useful to others here:
The free IEX feed returns extended-hours bars, and unevenly, one liquid name gave me 511 bars over 90 days where two others gave 438. I had to filter the replay timeline through the NYSE calendar to stop half my tickers sitting frozen at pre-market moments while a couple moved.
Bars are stamped at the start of their interval, which I got wrong at first. The hourly bar stamped 13:00 closes at 13:59, so my simulated clock was reading a full hour behind the price it was displaying.
Also worth knowing: adjustment=split matters for any continuous historical series. Raw prices across a split look like an overnight collapse that never happened. Fine when you’re reading one price at a time, not fine when you’re walking 90 days in sequence.
It’s free, Windows only, and it needs your own Alpaca keys plus an AI provider key: Anthropic, OpenAI, or a local model through Ollama. Running it costs real money in AI credit, including in sim, since every round is a live model call. Roughly a couple of cents a minute on a cloud model, nothing at all on Ollama. I’d rather say that up front than have someone find out by watching their credit drain.
No performance claims. It doesn’t predict anything and I’m not suggesting any of it is profitable. It’s just a really fun way to interface with your account and ultimately the characters do whatever you tell them.
Heads up that the installer isn’t code-signed yet, so Windows will throw a SmartScreen warning on first run. Certificate is in progress.
Happy to answer anything about the architecture, and genuinely interested in what breaks.