Yo, we are coooking! This is my testing deployment "ryto.pro" which will itself turn into a tool for AI agent use. But, the important thing is that I've had a GTM AI Agent process generate the following information. I've kept it vague (this is publicly listed after all) but I hope it sparks some imagination and answers some parts of my approach with local compute for safety. Chat soon :)
ryto.pro · a RelayPlex testing ground
For Matthew
How a locally-hosted AI agent can own its IP and still reach for scaled-up intelligence.
This domain is my personal testing ground for building with RelayPlex — a platform that gives you hosting, durable state, secrets, and auth as a few composable primitives. New site, live instantly, no DNS or SSL wait. Everything under ryto.pro is me pushing on what it can carry.
The most complex thing I'm carrying on it is Ryto: a long-polling contextual-intelligence tool that watches external websites change over time and lets AI agents reason about those changes efficiently, rather than re-reading the whole web on every query. I'm building it on RelayPlex as an example of a genuinely complex solution the platform can deliver — and it's the first thing that needs the scheduling-control primitive still in development. That primitive is what this page is really about.
Scale or secrecy — pick one?
Scaled-up intelligence work — big models, large research runs, fan-out job dispatch — wants to happen out in the cloud, where the compute is. But the most valuable context you have, your IP and your proprietary data, is exactly what you can't afford to ship out there. The usual setup forces a bad trade: keep your IP local and stay small, or go big and leak your crown jewels.
A marshaling service for a local agent
RelayPlex already ships the existing solution set — hosting, durable state (the published / submissions / workspace buckets), secrets in a vault used by reference, and edge-verified sessions. Add the scheduling-and-control primitive to that set and something new falls out:
A locally-hosted AI agent — running on your own node, where your IP lives — can use RelayPlex as a marshaling service for scaled-up intelligence. The agent stays home with the sensitive context; the scheduler is the governed doorway it dispatches outbound work through:
- Job dispatch — fan tasks out to scaled compute on a schedule, idempotent per run, with overlap, retry, and back-off handled by the platform rather than your code.
- Research — long-running, respectful pulls from external sources that return cleaned results, not raw dumps, back to the local agent.
- Anything bounded and scheduled — "go do governed, auditable work out there, and bring only the answer home."
The local agent keeps the IP; the scheduler marshals the scale. You own your crown jewels and still reach planetary-scale intelligence — because the sensitive context never leaves your node, only governed, scheduled requests do.
Control is the directionality, not a bolt-on
This is secure by construction, not by policy memo. Every primitive's direction does the security work: SSRF-guarded egress, per-host allowlists that default to deny, secrets injected server-side by reference so plaintext is never returned, append-only owner-read trails that make every dispatch auditable, and capability tokens that can write but never read back. Control isn't a feature you switch on — it's the shape of the platform.
Ryto is the first consumer
Ryto is the canonical case for marshaling: keep the reasoning and the history close, send only governed, scheduled pulls outward. It needs scheduled, respectful egress to watch the world change and reason over it — which is precisely why I'm building it as the forcing function for the scheduling-control primitive. Own the IP, borrow the scale. That's the bet.