I wanted an answer to one question: what language should I use if I’m building a Linux desktop GUI app with an AI coding assistant doing most of the work. Not “what’s the best GUI framework” in general — specifically, what works well when Claude is writing most of the code.

The setup

Opus is good at research but bad at staying in its lane. Give it one open question and it drifts into five related ones. So I split the research into three separate subagents, each with a hard scope boundary and an explicit “out of scope” list:

1. AI codegen fit — how well do LLMs actually generate and debug code for each language/framework combo
2. GUI framework landscape — native look, packaging (Flatpak, AppImage, .deb), maintenance, footprint
3. Community discourse — what people on Hacker News and the Qt forums actually say about doing this with AI

Same list of 8 stacks in all three: Python (PyQt6/PySide6, GTK4/PyGObject), Rust (egui, iced, Slint, gtk-rs), C++ (Qt, GTK), C#/.NET (Avalonia), Go (Fyne, Wails), TypeScript (Electron, Tauri), Dart (Flutter), Kotlin (Compose Multiplatform). Each agent got told what the other two were covering, so no overlap and no scope creep.

What came back

Codegen fit tracks language, not framework. Python resolves 52.2% of Multi-SWE-bench issues. C++ gets 14.7%, Rust 15.9%, Go 7.5%. The gap between languages is much bigger than any gap between frameworks in the same language.

Rust splits opinion hard. Some developers on Hacker News call it “nearly the perfect language for LLMs” — the compiler forces the agent to fix its own bugs before you even see them. Others report constant lifetime errors and the agent spamming .clone() to make the borrow checker shut up. The numbers back the second camp more: RustEvo² benchmark scores 32-38% on API-behavior tasks, and it drops further on anything after the model’s training cutoff.

Qt’s own community recommends Python for this, explicitly. A Qt forum moderator said it straight: “If vibe coding is your goal… Python should probably be your choice. Much bigger code base to train the LLMs on.” Qt has since shipped agent skills for exactly this reason.

C#/Avalonia has the best tooling, even if the language isn’t the top codegen performer. Avalonia ships a Build MCP built “so it stops hallucinating APIs,” plus a DevTools MCP that lets the agent inspect and screenshot the running app while it works. Smallest footprint on the list too — NativeAOT builds under 15 MB.

Only two stacks look native on Linux. GTK4 (GNOME) and Qt (KDE). Everything else — Electron, Tauri, Flutter, Compose Desktop, Rust GUI toolkits — draws its own widgets and looks foreign on both desktops.

What I’d pick

Python + PyQt6/PySide6. Best-documented framework pairing that exists, highest AI codegen success rate measured, and it’s the one the Qt community itself points to for AI-assisted work. Trade-off: heavier footprint, and it only looks native on KDE — use GTK4/PyGObject instead if GNOME integration matters more than raw AI throughput.

If I wanted the tightest edit-verify loop instead of the biggest training corpus, C#/Avalonia would be my second choice — the MCP tooling around it does a lot of the correction work Python doesn’t need in the first place.

I wouldn’t reach for Go or Electron/Tauri for this. Go has the weakest measured codegen fit and almost no published evidence either way. Electron/Tauri reports on Hacker News directly contradict each other — some people had their best results there, some had their worst, no pattern.

One honest gap: almost none of this discourse is Linux-specific. Most war stories are cross-platform or macOS. Worth a quick spike on your own setup before committing.