Short answer: a bring-your-own-model mode is coming to every Harpy installer. You’ll pick in the setup wizard whether Harpy uses the hosted path it uses today or a model running on your own hardware. The rest of the experience — the modules, the dashboard, the journal, the scheduled scans — stays the same. This post is a short note on what’s coming, what you’ll need to run it, and the trade-offs we want operators to understand before the alpha opens.
What’s coming
One extra question during setup: where should Harpy do its thinking? Two choices.
- Hosted (default, unchanged). What every Harpy installer does today. Fast, broad, strong on research, subscription-priced.
- Local (new). Harpy uses a model running on your own machine. Fully offline after setup. Nothing about your strategy, journal, or positions ever leaves your hardware.
Every other Harpy feature works identically between the two modes. The modules you’ve installed, the dashboard view, the paper / live toggle, the invite system, the 31-day lease — none of it changes. The one visible difference on the local path is a small badge in the dashboard header so you always know which mode a given scan ran in.
Why we’re building it
The most common question we get from invite recipients is some version of “can I run this without calling an outside service at all?” Three reasons tend to come up, in this rough order:
- Ongoing cost. Subscriptions add up, especially on the 24/7 modules.
- Privacy. Some operators want their thesis and position sizing to stay entirely on their own machine — not for any single reason, just as a default posture.
- Consistency. Harpy is already local-first in every other way: credentials on disk, dashboard on localhost, scans firing on your own scheduler. The model was the one piece still reaching outside. Closing that gap makes the product internally consistent.
What you’ll need
A capable machine
Realistic floor: a modern gaming GPU with 24 GB of video memory, or an Apple Silicon machine with 64 GB of unified memory. If your machine is below that line, stay on the hosted path — local inference will be too slow to keep up with the scanner’s cadence, and the Harpy wizard will say so plainly if it detects unsuitable hardware.
A one-time install step
You’ll install a small local model runner (same category of tool that open-source AI communities already use heavily). It’s a well-documented, ~15-minute install with a first-class Windows installer. The Harpy wizard links directly to it and checks for its presence before the local path becomes selectable.
A model from our recommended list
Harpy will ship with a short, vetted list of recommended models to choose from — so you don’t have to be a model-picker to get a working setup. If you’d rather swap in a different model you already have, the wizard exposes that as an advanced option.
The honest trade-offs
Research quality
Hosted Claude is the quality ceiling, and we’re not going to pretend otherwise. On deep research — reading a Fed statement and inferring the macro regime flip, say — local models are a clear step down. On structured rubric work — applying a confidence floor, generating a target list — the gap is smaller but still real. Operators on the local path should expect more rejections and tighter shortlists. The dashboard will surface this: when the local model doubts a target, you’ll see the reason written into the journal, not a silently-dropped pick.
Speed
A full scan pass on a capable local machine takes meaningfully longer than the same pass on the hosted path. It still fits inside every scheduled cadence comfortably, except the very fastest one (the 5-minute Max+ position monitor). If you’re on that cadence and you flip to local, Harpy will automatically skip any cycle that hasn’t finished before the next one is due, and log it.
What stays exactly the same
Live trading. Orders still route to your broker over a real internet connection exactly as they do today — Alpaca to Alpaca, Kraken to Kraken, and so on. The only thing that moves to your machine is the research step. Your keys, your journal, your position sizing — all unchanged, all stored exactly where they are today.
Shipping order
- Paper first. The alpha will open inside the Paper installer — no real money, every module unlocked. You can watch the local path produce targets across all seven markets for a week before anything touches live.
- Trade and Crypto next. Once the paper alpha is stable, the same toggle lands in Trade and Crypto — the two modules operators spend the most time in day-to-day.
- Everything else after that, in the order operators actually ask for it. Polymarket, Kalshi, Horses, and Greyhounds all get the toggle.
The strategy has always been yours. Now the model can be too.
How to get on the alpha list
If you’re an existing invite holder and you want in, just reply to the invite email with three lines: the hardware you’d run it on, which modules you actually use, and whether cost or privacy is driving the interest. That’s it — no form, no survey. If you’re not an invite holder yet, leave your email on the apply page and we’ll pull from the waitlist when the alpha opens.
We’ll publish a short follow-up the week it goes live. Nothing above is locked — the post exists so the plan is public and operators can push back before we build the wrong thing.