What if you could describe an app in plain English and get a real, native desktop application — not a web preview, not a sandbox, but an actual compiled program running on your machine?
That's what I've been building for the past months. It's called N.E.O. (Native Executable Orchestrator), and it's now open source.
The idea: Everyone knows Claude Artifacts and ChatGPT Canvas. They're great for web previews. But they run in a browser. I wanted to know: why doesn't something like this exist for native desktop apps? So I built it to figure out where it breaks.
N.E.O. takes your prompt, sends it to an AI (Claude, ChatGPT, Gemini, or a local model like Ollama), gets back C# code, compiles it at runtime using Roslyn, and displays the result in a child process. No IDE or SDKs needed. No project setup. Just describe what you want.
Here's a prompt I use as a showcase:
"Build an app that captures the system's audio output in real time via WASAPI and generates a spectacular live animation for a techno club. The animation should react fluidly to frequency, volume, bass, and beats."
How it works under the hood
You type a prompt
→ AI generates C# code
→ Roslyn compiles it to a DLL (in memory)
→ DLL is streamed via Named Pipes to a child process
→ Your app appears on screen
Each iteration builds on the previous one. When you ask for changes, the AI sends a unified diff patch — no full rewrite needed. If something doesn't compile, N.E.O. automatically feeds the error back to the AI and retries.
The child process runs in an optional Windows AppContainer sandbox, so generated code can't mess with your system.
What it supports
- UI frameworks: WPF, Avalonia, React (via WebView2)
- Export: Standalone executables for Windows (WPF), or Windows/Linux/macOS (Avalonia)
- AI providers: Claude, ChatGPT, Gemini, Ollama, LM Studio
- Built-in AI capabilities: Image generation, image analysis, speech-to-text, text-to-speech
- Editing: Visual designer mode (click-to-edit), branching undo/redo history
- Extras: NuGet package resolution at runtime, optional Python integration
The tech stack
- .NET 9/10, fully self-contained (no external SDK dependencies)
- Roslyn for runtime compilation
- Custom IPC protocol over Named Pipes (framed binary protocol with 32-byte headers)
- Collectible AssemblyLoadContext for plugin isolation
- Embedded AppHost templates for cross-platform export
Why I'm posting this
This started as a hobby project — I was genuinely curious whether this approach could work at all. Turns out it can.
But building something alone only gets you so far. I'm looking for people who find this interesting and want to contribute — whether that's new features, bug fixes, testing, or just honest feedback about what's broken.
The project is pre-alpha. It works, but there's plenty to improve.
Links
GitHub: github.com/hf75/N.E.O
Demo video: youtube.com/watch?v=6OZxm7ZEVU0
License: MIT
Top comments (0)