GUIOS is a new kind of software: the agent-native app — an interactive application whose interface, state, and behavior are generated and maintained by an agent. Describe what you need on a whiteboard, and it becomes a clickable, fillable app you can keep using — where every action genuinely drives the services it connects to.
The local demo runs at http://127.0.0.1:8787/ · this homepage runs at http://127.0.0.1:8090/
Research preview · Every screenshot on this page comes from the actual running system — not a mockup.
Real screenshot · Every card on the whiteboard is a live app generated from a single sentence.
Not a mockup, and not a stitched-together conversation. GUIOS compiles your request into a validatable app spec, renders it instantly through a universal runtime, and hands execution to the app’s own agent backend.
Describe it in natural language in the input bar at the bottom of the whiteboard. Creating, modifying, connecting, querying — all through the same entry point.
A team of agents breaks down the request, designs the interface, defines actions and data, and picks real services — producing an app spec that can be validated, diffed, and rolled back.
The universal runtime renders the spec into an interactive interface. You click buttons and fill forms; the backend agent actually calls the APIs it connects to and writes results back into the state you can see.
Real screenshot · Open a card and you get a complete application interface.
A chatbot hands you a paragraph and leaves the actual work for you to do somewhere else. GUIOS gives you an interface you can click and use — the action completes right there.
What’s generated isn’t a picture you can look at but not click. Every button is wired to a real service — press it, and something actually happens.
State isn’t conversation history hidden inside a model. It’s visible, inspectable, durable application state — and it’s still there when you come back.
An agent-native app has to honor five contracts at once. They turn the definition — interface, state, and behavior maintained by an agent — into promises you can check one by one.
What you’re doing and how far you’ve gotten is visible in the interface — this is “interface maintained by the agent.”
Every action you take dispatches into the app’s own agent workflow — never a static control or a dead handler.
Actions genuinely invoke the real APIs they name and create real entities — not simulated results.
What really happened is written back into the interface you can see — this is “state maintained by the agent.”
Close the app and open it again: your content and progress remain. State is persistent, not one-shot.
These five contracts are the foundation of LiveUIBench: every score is read from behavior that actually executed.
The whiteboard is already running on your machine. Say one sentence and watch it become a working app on the spot.
GUIOS is a research preview. Every screenshot on this page comes from the actual running system. To learn more or try it out, get in touch with the author.
