If you're interested in a GPU-rendered Rust UI framework, then maybe skip this vibe coded project, and check out some of following established projects:
List is hand-picked list, ordered by crates.io downloads. I had an LLM fetch the download counts and links. Webview based UI like Tauri and Dioxus Desktop are excluded. As are bindings to native toolkits and GTK.
IMO, the top GUI framework crates for Rust are windows or native-windows-gui, objc2 or icrate for MacOS and gtk4-rs for Linux only. Actually for MacOS / iOS I'd use uniffi to embed the rust core in a SwiftUI app.
In 2026 IMO it's much better to use an LLM to keep 3 separate native front ends in sync than it is to try and use any sort of cross platform UI.
Keeping 3 separate native front ends in sync with custom visual components means you end up building a cross-platform UI. The only apps you can build without custom components are trivial ones.
All 3 have a way of embedding webviews. In my perhaps limited experience, all desire for custom visual components are either specious or for read-only presentation views. In my opinion all controls should be native, but if you ask for a fancy data presentation I have no qualms reaching for a webview.
So long as the web views provide for accessibility (and thus UI automation), this sounds perfectly reasonable for the majority of applications.
Speaking of which, my dream is an "HTML light" standard with a subset of functionality specifically designed for local UI use, ideally with lightweight non-browser engines supporting this standard shipping with all major OS platforms as shared system components.
We had browsers that ran on systems with 8 MB RAM, after all, and while it's not a fair comparison, there's no technical reason why a system with three orders of magnitude more RAM should struggle to run more than a handful of applications merely because they use HTML and JavaScript for rendering.
I am building such a lightweight browser engine: https://github.com/dioxuslabs/blitz. It targets most of HTML/CSS, but leaves out (or makes optinal) a lot of the extra stuff that browsers pull in). It also has the option to use Rust instead of JavaScript for the scripting layer.
Though 8 MB RAM is a non-starter when a single RGBA texture @4k is ~30Mb and modern system often require 2 or 3 for buffering. Somewhere around the 60MB to 100MB mark seems to be the base for modern setups, at least on macOS.
https://crates.io/crates/dioxus-native was released 3 months ago and does not have a link to GitHub, I do not think it can be considered as established project yet.
(I am the maintainer of dioxus native, and also posted the comment above)
So:
- The initial version was released over a year ago (3 month is the most recent patch release of the stable version).
- It is part of the Dioxus project (https://github.com/dioxuslabs/dioxus) which has been going since 2021 (but the other renderers are web/webview based).
- The native renderer (https://github.com/dioxuslabs/blitz) has been under development since December 2024 (building a whole new web engine is just a big project).
You're ultimately kind of right that it's not as established as 1-4 in the list. But I think it justifies it's place with the other frameworks lower down the list as one of the lesser-known frameworks that compete technically with those further up the list.
Yeah, have you tried to do anything in egui? I tried today a ComboBox, but I wanted one with filter input box, "X" icon to clear it, and automatic focus.
I've given up when the code reached 300 lines and was still buggy/not working.
egui is not mature. I'd prefer declarative approach, data binding, and styling. Therefore will skip egui.
i've built pretty complex stuff in egui. any modern agent will one-shot that component, so i'd say it's more just a lack of familiarity with the library.
Slint is proprietary. Given that AI makes it so easy to spec out and build even entire GUI frameworks, why would you even go with proprietary anymore? And why doesn't the article showcase a full-fledged set of UI components?
I find this line pretty sad actually, what if someone wants to contribute code, written by hand? What if some beginner finds this project interesting and wants to invest own time into it?
Or human is completely removed from the development whatsoever..
> I find this line pretty sad actually, what if someone wants to contribute code, written by hand? What if some beginner finds this project interesting and wants to invest own time into it?
Then they would be better off contributing to a project that welcomes them.
> what if someone wants to contribute code, written by hand?
It's up to the maintainer. Sometimes it's just easier to listen to what's needed and do it the way you know and are efficient with than to review someone else's code.
> What if some beginner finds this project interesting and wants to invest own time into it?
There's nothing stopping them. They can even publish it in a fork.
Yep. There's other projects that have a "we want your ideas, but not your code" policy (like SQLite and Lua), but the "just tell your agent to tell our agent what to do" framing makes me depressed.
I did the same for a vibecoded project [0] but I would never do this for a human-written project. Simply put, if most contributions are similarly vibecoded and/or they will be reviewed/treated as such, then how is a PR/review from an unaffiliated agent any better than one managed by the project curator. Of course that human curation still needs to occur.
This is a natural progression (only on these types of projects). We accepted contributions because it actually helped and they were difficult and time consuming. When that is no longer the case, contributions are a burden, the actual problem/idea is the real value, hence the request to open the issue.
My main concern with this would be that for a library in particular I need a stable API and behavior (Hyrum's law). I do not know to what extent a fully autonomous agentic approach can ensure that. Note: I don't see a problem with preferring issues over pull requests (many of which would be poorly coded slop), just fully automating the handling.
Also a bit wild to assume that no person will just create an issue manually. This is not something I'd normally use an agent for.
It sounds like a black hat penetration tester's nirvana. All you have to do is open the right issue to trick the model into installing the right security vulnerability or backdoor, or better yet, a series of small security vulnerabilities that can be chained to create a tailor-made exploit. It's practically a hacker's choose-your-own adventure.
I suspect they'd be manually filtering which issues/features their agents attempt to fix/implement. Otherwise you end up with all sorts of features that simply make no sense for the project.
The approach is at least better than accepting slop PRs that could have backdoors injected into them.
Headless part is neat, but agents can already drive apps and take screenshots. As a side note, I have agent driving android tv using wireless adb and hdmi capture card + mpv rpc, its fun to watch.
Looks interesting, especially the headless part. Been building various apps lately for ios/ipados/macos for personal use and I would love something like this that would be truly cross platform. Xcode with agent use has been really nice to work with, and t3.code support for emulators has been amazing to use.
What is benefit of GPU-rendered UI? In regular 2d apps where I can select enable/disable GPU rendering, disabling doesn't make app feel slower, but it drastically reduces RAM and VRAM usage (VRAM is the one I care about, as I run local models). Why so many UI frameworks brag that they are GPU-rendered?
Other people have different hardware and different goals. A year or so ago I worked on a Linux laptop that was dog-slow. If I didn't enable GPU-rendering for the UI, it competed for the very limited CPU.
Using less power is generally the main benefit. On lower end hardware you may also notice a performance difference, depending on what the UI is trying to display.
When I made my ow GPU-renderered framework I found that it was possible to render the entire UI in one call/shader to the UI. All primitives and styles can be passed as a updatable data blob that the single GPU shader handles in one call. Do most of the GPU-rendered frameworks do it this way?
It has some tradeoffs to consider but seems quite good overall.
An approach I've found useful, albeit for a DSL that ships a binary for execution, is to give a way for the LLM agent to get a high-level topic list so it can drill down all the way to specific language/API implementation details.
Perhaps with this project the author could make a website that does the same?
Sorry, what? I feel like I've studied the Rust UI ecosystem pretty well, and the examples look nice, but if I've learned anything it's that "native" and "GPU-rendered" are oxymoronic.
Likely the fonts will be fuzzy on various platforms, accessibility will be under-baked and it will land in the bucket with every other UI framework depending on winit. I like the personalities at anomaly, but I'm quite confused about what's even driving a decision to build something like this versus contributing to egui, GPUI, Iced, etc.
It's definitely possible to do sharp fonts in the Rust GUI ecosystem (although you certainly won't get it out of the box). With the native text libraries if you want to (both Servo, Zed/GPUI, and anything using Skia do this), or emulated with Skrifa + Vello.
Yeah. Presumably "native" here just means "not Electron", which is the sales pitch for most AI-generated Rust libraries these days. ("Tell your agent to rewrite your Electron app into Rust~! :rocket emoji:")
It would be fun to port some apps over to this, I'm worried at the speed it moves, It would be hard to settle on a commit. I'd rather fast updates over slow, but I feel like I'd have very little trust on this between updates, especially how many things could be broken after updating.
I have the same annoyance. It's continually maddening to see "native" plastered everywhere only for it to be some non-native junk. I miss when native meant something like Qt.
Qt is an interesting case, because it doesn't use the native platform's widget code*, but does a good job of looking and behaving as though it does. That's usually close enough for me.
Various other toolkits have tried this approach, but unfortunately, I'm not aware of any others that succeed. Common deficiencies include keyboard navigation and text rendering.
*(Except on Qt-based desktops like KDE Plasma, of course.)
I've been thinking a lot about the term "native" as I come from mobile. When writing a core in Rust (which I like to do), and then layering on the operating system's UI SDK components per platform (which I also like to do), what is native and what isn't, especially when you throw in React Native to the discussion?
I like your definition quite a bit, but I need another word for, "code that doesn't need a runtime to execute."
I ended up using "experience layer" to describe the SDK components.
Changing times. Nowadays a lot of software UIs are developed by people with a web UI background. Those people won’t learn the OS-native UIs because that would mean they’d give up on their existing knowledge. Because of that native means “not running in an embedded web browser” to many nowadays.
Native often means "not a wrapper around a webview".
If you define it as "using the widgets provided by the operating system", then "native" is by definition incompatible with "cross-platform." And what does that mean on Linux, where the operating system doesn't provide widgets? Are gtk and Qt not native?
When discussing GUI toolkits, "native" should be understood to mean "native to the host's graphical environment".
"Native to the OS" might or might not be equivalent, depending on the OS and what one considers "OS" to mean. (Does it mean a kernel? A desktop distro? Something else?) Because of this ambiguity, it's not a very practical way to approach the discussion. Different people attach different meanings to that phrase.
Some frameworks are native to Qt. Others are native to Gtk. Practically none are native to Linux. ;)
I don't think "cross-platform" is incompatible with "native". It has been achieved more than once using toolkit wrappers, usually selected at build time.
I was thinking "Builds and tests with a fast-moving Rust code base like this must be painful", then I noticed the total absence of tests: https://github.com/anomalyco/rune/commits/dev/
You're not shifting left unless you guarantee local tests always run, faithfully to production. That rarely happens. Local testing should be supplementary. We use https://hk.jdx.dev/
We do ensure this, requiring signed proof, running in containers that would be identical to github runners (more or less). Now, this could be spoofed by an engineer, but we operate in a high trust environment
This is what a vibe coded UI framework looks like.
Have a read of the clean documentation that they never wrote or read. [0]
All the clean code, architecture principles thrown out of the window in favour of hundreds of code smells, anti-patterns, zero tests, and tons of unmaintainable slop.
Damn, this README is more of a DONTREADME. Even the root README is for agents, not people. The more I see stuff like this, the more I think about becoming a welder or a carpenter...
The site is wild, but the little triangles in the bottom right corners of each window don't work properly on chrome desktop in windows 11. They aren't clickable, you can only resize the windows horizontally. Which seems weird considering how polished the rest of it looks.
It's a fun tech demo, egui isn't meant to make websites. It's primarily for integration into games/other real-time-3D-rendered applications, and secondarily for GUI applications via eframe.
But I agree, I think they would be better served by making a normal website and linking to the demo prominently (For comparison, see iced's website https://iced.rs, which is great IMO)
In 2026 IMO it's much better to use an LLM to keep 3 separate native front ends in sync than it is to try and use any sort of cross platform UI.
Speaking of which, my dream is an "HTML light" standard with a subset of functionality specifically designed for local UI use, ideally with lightweight non-browser engines supporting this standard shipping with all major OS platforms as shared system components.
We had browsers that ran on systems with 8 MB RAM, after all, and while it's not a fair comparison, there's no technical reason why a system with three orders of magnitude more RAM should struggle to run more than a handful of applications merely because they use HTML and JavaScript for rendering.
Though 8 MB RAM is a non-starter when a single RGBA texture @4k is ~30Mb and modern system often require 2 or 3 for buffering. Somewhere around the 60MB to 100MB mark seems to be the base for modern setups, at least on macOS.
New rule: Programmers with limited UI experience (that includes myself) should abstain from weighing in on what UI designers and programmers need.
So:
- The initial version was released over a year ago (3 month is the most recent patch release of the stable version).
- It is part of the Dioxus project (https://github.com/dioxuslabs/dioxus) which has been going since 2021 (but the other renderers are web/webview based).
- The native renderer (https://github.com/dioxuslabs/blitz) has been under development since December 2024 (building a whole new web engine is just a big project).
You're ultimately kind of right that it's not as established as 1-4 in the list. But I think it justifies it's place with the other frameworks lower down the list as one of the lesser-known frameworks that compete technically with those further up the list.
I've given up when the code reached 300 lines and was still buggy/not working.
egui is not mature. I'd prefer declarative approach, data binding, and styling. Therefore will skip egui.
Or human is completely removed from the development whatsoever..
Then they would be better off contributing to a project that welcomes them.
You could fork and reference it in the issue if you still wanted to contribute.
It's up to the maintainer. Sometimes it's just easier to listen to what's needed and do it the way you know and are efficient with than to review someone else's code.
> What if some beginner finds this project interesting and wants to invest own time into it?
There's nothing stopping them. They can even publish it in a fork.
This is a natural progression (only on these types of projects). We accepted contributions because it actually helped and they were difficult and time consuming. When that is no longer the case, contributions are a burden, the actual problem/idea is the real value, hence the request to open the issue.
0 - https://github.com/cretz/vitaslop
Of course, one could conclude that another agents screwed up before, still was a good experience.
Also a bit wild to assume that no person will just create an issue manually. This is not something I'd normally use an agent for.
Which gate is easier to maintain?
1. Check issues for malicious intent
2. Check PR for malicious intent.
I think in PR requests it is easier to introduce purposeful bugs or obfuscated code than it is to hide them in issues.
So I think gating on issues is easier and more secure.
The approach is at least better than accepting slop PRs that could have backdoors injected into them.
https://lovro-bikic.github.io/software-line-specialist/
This is cool. I hope SwiftUI / UIKit borrow this idea for native mac apps.
- Accessibility? E.g. Screen reader support?
- Alternative inputs? Full keyboard support?
- Full internationalization, including right-to-left and diacritics?
What happens when no GPU is available?
Just curious as this matters for it to be a serious library and not a proof of concept.
It has some tradeoffs to consider but seems quite good overall.
An approach I've found useful, albeit for a DSL that ships a binary for execution, is to give a way for the LLM agent to get a high-level topic list so it can drill down all the way to specific language/API implementation details.
Perhaps with this project the author could make a website that does the same?
The double-edged sword of choice!
Sorry, what? I feel like I've studied the Rust UI ecosystem pretty well, and the examples look nice, but if I've learned anything it's that "native" and "GPU-rendered" are oxymoronic.
Likely the fonts will be fuzzy on various platforms, accessibility will be under-baked and it will land in the bucket with every other UI framework depending on winit. I like the personalities at anomaly, but I'm quite confused about what's even driving a decision to build something like this versus contributing to egui, GPUI, Iced, etc.
But I agree with the reset
Various other toolkits have tried this approach, but unfortunately, I'm not aware of any others that succeed. Common deficiencies include keyboard navigation and text rendering.
*(Except on Qt-based desktops like KDE Plasma, of course.)
I like your definition quite a bit, but I need another word for, "code that doesn't need a runtime to execute."
I ended up using "experience layer" to describe the SDK components.
If you define it as "using the widgets provided by the operating system", then "native" is by definition incompatible with "cross-platform." And what does that mean on Linux, where the operating system doesn't provide widgets? Are gtk and Qt not native?
When discussing GUI toolkits, "native" should be understood to mean "native to the host's graphical environment".
"Native to the OS" might or might not be equivalent, depending on the OS and what one considers "OS" to mean. (Does it mean a kernel? A desktop distro? Something else?) Because of this ambiguity, it's not a very practical way to approach the discussion. Different people attach different meanings to that phrase.
Some frameworks are native to Qt. Others are native to Gtk. Practically none are native to Linux. ;)
I don't think "cross-platform" is incompatible with "native". It has been achieved more than once using toolkit wrappers, usually selected at build time.
[0]: https://rune.build/
e.g. https://github.com/anomalyco/rune/tree/dev/crates/rune-kit/t...
https://redbadger.github.io/crux
Have a read of the clean documentation that they never wrote or read. [0]
All the clean code, architecture principles thrown out of the window in favour of hundreds of code smells, anti-patterns, zero tests, and tons of unmaintainable slop.
[0] https://github.com/anomalyco/rune/blob/dev/crates/rune/READM...
But I agree, I think they would be better served by making a normal website and linking to the demo prominently (For comparison, see iced's website https://iced.rs, which is great IMO)
This isn't much better. It's super laggy with high CPU usage.