From Agent
to App

Agent thinks. App appears.

Three steps from natural language intent to a fully interactive application, rendered natively on any platform.

Not just chat.
Full applications.

Taking inspiration from the A2UI protocol and running on Makepad 2.0, A2App turns AI agents into app generators — secure, declarative, and universal.

From prompt to app,
in seconds.

Watch an AI agent transform a simple request into a fully interactive application.

One agent response.
Every screen.

The same declarative UI renders natively on desktop, tablet, and mobile — using each platform's own design system and components.

What will your agent build?

From internal tools to customer-facing products — agents can generate any interface on demand.

Real-world examples

Projects bringing A2App's ideas to life — from Makepad 2.0 native apps to prompt-injection-resistant agent runtimes.

Matrix Client

Robrix

A multi-platform Matrix chat client written in Rust, exemplifying Project Robius with Makepad 2.0. One codebase ships natively to macOS, Windows, Linux, iOS, and Android.

Rust Makepad 2.0 Project Robius

View on GitHub →

🔒 lsp_agent

Agent Runtime

lsp_agent

A prompt-injection-resistant agent runtime. A VS Code extension where an AI agent (GitHub Copilot) writes and launches web applications into a webview — with all inference and file access gatekept by a privileged LSP server, and Automerge coordinating isolated processes. This is the A2App pattern in practice: agents describe the app, a hardened runtime enforces the rules.

Rust LSP Automerge Injection-Resistant

View on GitHub →

Constraints that actually constrain.

When AI agents generate interfaces, the runtime does the hard work. Rust's guarantees and Makepad 2.0's declarative model give agents expressive freedom — without giving up the rules.

The practical question is not whether to allow AI agents in markets but how to make them work well. Formal verification offers something concrete: constraints that actually constrain, rules that cannot be silently violated, guarantees that hold regardless of what the model learned.

If these trends continue, we may be witnessing another industrial revolution. Before mechanization, most work in the physics sense of force times distance was done by muscles. The steam engine and its descendants changed that. Most thinking used to be done by brains. In the near or far future, that may change too. This is not inevitable. It presumes breakthroughs that are plausible but not guaranteed. But the trajectory is clear enough to take seriously.

— Stephen Diehl, Zero to QED, Ch. 22 "What Comes Next"

This is the bet behind A2App + Makepad 2.0: agents describe what the UI should be; a Rust-native, declarative runtime enforces how it behaves. Expressive freedom for the model, hard guarantees for the user — apps that cannot be silently violated, no matter what the agent learned.

Trust what you render.

A2App uses webviews to embed rich content — and when an agent can influence what appears inside that surface, "probably safe" isn't good enough. Prompt injection, unexpected navigation, and tampered DOM state become real attack surfaces. The engine itself needs guarantees.

Inspired by A2UI.
Built with Makepad 2.0.

A2App draws inspiration from Google's open A2UI protocol for declarative, agent-driven interfaces — and brings those ideas to life with Makepad 2.0, a Rust-native UI runtime that renders across desktop, mobile, and web.