Short answer: Local-first wins on key custody, data residency, audit transparency, recurring cost, and long-tail flexibility. Cloud-hosted wins on uptime, colocation latency, zero-config onboarding, and cross-device continuity. The right choice depends on your strategy cadence and which threats keep you awake at night. For a research-driven retail operator on a 5-minute-or-longer cadence, local-first is structurally better; for a colocated HFT shop, it’s a non-starter.
The two architectures
A cloud-hosted trading terminal runs the scanner, the model, the execution engine, and the dashboard on servers operated by the vendor. You log in through a browser, you give the vendor your broker keys (or oauth into a brokerage they have a partnership with), and you receive signals and execute trades through their UI. Examples: most TradingView alert-based bots, Composer, hosted QuantConnect deployments, most fintech “copilot” products.
A local-first trading terminal runs the scanner, the model invocation, the execution engine, and the dashboard on your own hardware. You install software, paste credentials into a wizard, and the application registers scheduled tasks that fire on your machine. The vendor never sees your keys, never sees your trades, and never sees your journal. Examples: Harpy, self-hosted Backtrader, Hummingbot (decentralised side), most home-grown Python scripts running under cron.
Where local-first wins
Key custody
The single biggest structural difference. A local-first terminal stores your API keys in a per-user directory on your disk; a cloud terminal stores them in the vendor’s database (usually encrypted, but readable to the vendor under a court order, an insider attack, or a breach). For high-permission keys — withdraw rights on a crypto exchange, full margin on a stock account — the difference is material.
Data residency and audit
Every scan’s input, output, and rationale is a file on your disk. You can grep your own decisions, you can diff today’s targets against yesterday’s, you can reproduce any single decision the agent made by replaying the skill doc against the cached input. Cloud terminals generally expose only the surface metrics (P&L, position history) and not the agent’s working memory.
Recurring cost
A local-first terminal pays for itself in compute already on the desk and a flat-rate model subscription. A cloud terminal adds a SaaS line item that scales with seats, account count, or trading volume. Over five years, the cumulative SaaS bill on a cloud terminal will substantially exceed a one-time hardware purchase plus a flat subscription.
Long-tail flexibility
A cloud terminal exposes the strategies its vendor has shipped. A local-first terminal where the strategy lives in a plain-text file (as in Harpy’s skill documents) lets you edit the strategy on disk and run your version on the next cycle. No feature request, no roadmap dependency, no three-quarter wait.
Where cloud-hosted wins
Uptime
A cloud terminal runs in a datacentre with redundant power, redundant networking, and a paid SRE. Your home machine runs on a consumer ISP, a single power feed, and you. If your electricity drops at 2:47 a.m. on a Sunday, the cloud scanner’s next cycle still fires; yours misses.
Colocation latency
For strategies whose edge is measured in milliseconds — pure arbitrage, market-making, latency-sensitive momentum — a node colocated next to the exchange wins. A home machine on residential bandwidth cannot compete. This is why genuine HFT is structurally a cloud problem.
Zero-config onboarding
Cloud terminals can be live in minutes — sign up, oauth in, go. Local-first terminals require you to install software, paste keys into a wizard, register scheduled tasks, and confirm the first cycle ran. The on-ramp is steeper.
Cross-device continuity
A cloud terminal is the same instance whether you check it from a phone or a laptop. A local-first terminal lives on one machine; if you want to read the dashboard from somewhere else, you have to expose the local server (Tailscale, self-hosted reverse proxy, etc.) or accept that your dashboard is desk-bound.
How to decide
Three questions sort most of the choice cleanly:
- What’s your strategy cadence? Sub-second: cloud. Minutes-to-hours: local-first is fine. Daily or weekly: local-first is preferred.
- What permissions do your API keys carry? Trade-only with no withdraw permissions: either is fine. Full custody including withdraw: local-first is the safer default.
- Will you actually maintain a home machine? If the answer is “no, I want to forget it exists,” cloud wins by default — local-first rewards operators who will keep the host healthy.
Where Harpy sits
Harpy is committedly local-first. Every scan runs on your machine, every credential lives in your AppData, and the license check itself is offline-validated against a cryptographic 31-day lease. We made this choice deliberately and we’re honest about its costs: Harpy is not for operators who want a phone-first dashboard or who can’t keep a machine awake. It is for operators who want a system they can audit, edit, and unplug.