I had a look through the code to see how it manages not to use a pile of C libraries with cgo like other Go GUI libraries.
The answer is platform dependent:
Windows loads the relevant DLLs by hand and calls them. This is a well established technique in Go programs and due to the super stable DLL interface works well.
Linux has an x11 and Wayland backends and these implement (through a library) the wire protocols directly in Go which is nice and will make cross compilation and distribution easy.
macOS does appear to use cgo to access the cocoa libraries. macOS doesn't like statically linked Go programs anyway though as they don't use system name resolution so this isn't a bad compromise, but will mean macOS stuff needs to be built on macOS I think.
I didn't see Android or iOS support.
A nice innovative approach to GUI building. Since the lowest common denominator for the backends is an RGBA buffer, this will bypass all accessibility things the OS provides.
The above gleaned after a few minutes reading the source so may not be 100% accurate.
Theres is another new project which does NOT rely on CGO https://github.com/gogpu/ui Its basically the whole WebGPU implemented in Go. Theres also Gio, but not sure if its still active. CGO is a huge pain…
> Mobile is under consideration (no decision yet). If it is supported, it will be limited to utility-style apps — not games or rich multi-touch experiences.
No Multiple Windows so even desktop apps will be limited to "utility-style apps".
I compiled and ran the process_monitor example on linux: it works, compiles fast and is about 10mb. Also cross-built for windows and it's 8.4mb. Can't build for macos/arm64
(Under wine the windows exe doesn't render text. weird.)
it's very hard. These “change the entire world” commits make for a history that is impractical to follow for a human, and therefore of little interest to me.
I think a challenge for a vibe coded library to gain adoption is if there is a lack of human time investment in its creation, how do we know there will be investment in its maintenance?
Most open source projects are abandoned because the author never finds enough time to work on it or can't muster the focus and attention needed.
It is not true that there is "lack" of human investment in the creation of this. If anything, I spent the last two weeks glued to the screen most of the time, to a degree I have never experienced before, building out all the different areas that lead to this release.
I will not mention the monetary investment because it's not the type that matters here.
Attention, which is arguably the most scarce form of investment, was invested in ample amounts.
> I spent the last two weeks glued to the screen most of the time, to a degree I have never experienced before,
With respect, a single period of manic hyperfocus is not what makes a project, it is being in it for the long-haul and putting in the hours here and there even when it isn't fun any more and you have many other things you might rather be doing.
To be clear I don't want to detract from what you've made here. Agentic engineering done right does still require a lot time and attention IMO. But you've still got a challenge on your hands to demonstrate to potential adopters that you're in it for the long haul.
but if you other ppl to read, use or contribute, they will probably need the "work history" more than a "publish history". because nobody is going to read those big ass commits, specially when unsure about it worthiness.
ok, so because of that you manually export changes you want to publish to a new repo? i'd start separating the repos. personally i would be uncomfortable using a repo that the developer does not use for their own development. if you ever want to accept contributions this might become an issue. but that's just my feeling.
Source is the new binary. Specs and requirements are where the engineering is happening now. Of course, this was anticipated in 1971 with the creation of the very successful PRIDE methodology, but it's taken a few decades, and a complete devaluation of their profession due to AI, for programmers to catch up.
This is the PRIDE I meant. It's hard to tell if MBA (more recently, M&JB Investment Company) is still a going concern. Milt Bryce's son Tim was keeping the lights on but he died of cancer at the end of 2023.
> Experience has shown us that an immediate mode API is the only sane way to program GUI applications
I wonder how long till they pivot away from this belief. I feel like everyone in UI goes through this phase as some point, but in the end it doesn't scale to truly complicated UI
Yeah. Based on my personal experience I think some kind of hybrid of old-school imperative retained and declarative retained, both with granular reactivity is probably the correct balance for "serious" high-utility desktop applications. Declarative approaches are great for smaller components but become a nightmare for anything much more complex than a relatively simple mobile app while imperative requires a lot of extra legwork at the component level, and as I understand (which may be incorrect) immediate mode makes certain types of optimization more difficult.
In my experience, it's the opposite. Immediate mode GUI, or at least a functional and declarative approach, is the only way I've seen it scale well. It's more modular and scale-independent. On the other hand, retained mode, or imperative/OOP approach to state management, becomes complicated and monstrous quickly; it's the dominant style and can be made to work OK, but typically hellish to maintain or develop beyond a certain scale.
Admittedly I'm simplifying too much and conflating paradigms. My preference is something like "functional core, imperative shell" or maybe "immediate-mode core, retained-mode shell" if that makes sense.
Obviously most GUIs are not nearly that complex so immediate mode can get you quite far. Its biggest limitation is that it makes it hard to do some layouts. Your GUI layout becomes dictated by your data dependencies which is quite awkward.
Absolutely zero difficulty redoing this in a react style renderer. The only complexity is being careful with your data dependencies so as to not needlessly rerender.
Each pane is easily isolated, can share data with a view model scoped properly, etc. Writing it in an imperative toolkit is a "oops I forgot to update my data here" kind of hell. Data binding makes it slightly less worse
> Absolutely zero difficulty redoing this in a react style renderer
I would not classify a react style renderer as "immediate mode". It has aesthetic similarities with IM GUIs, but ultimately there is a fully retained tree under the hood, that gets diffed/mutated on every update
What does "scale"even mean in UI context? 10 or 100 controls in app makes difference how exactly? Retained apps redraw when needed, they are idle most of the times. How redrawing every frame helps to scale?
Does "immediate mode" mean the UI refreshes at a constant 60fps rate like a video game, redrawing everything all the time?
No.
It means that you build the UI by describing what it should look like now, based on the data / state you own, without referencing any existing "widget" object or trying to manipulate it.
Scale is not about the number of buttons, but the structure of the data.
You have a list of objects, within each objects you have several fields, some of them lists, some of them maps. within some of those sub-items you have other lists and maps, nested arbitrarily.
This would be hell to manage for a retained mode UI. You have to mirror the application data into a widget tree and keep all the elements in sync, all the way down to the arbitrary depths of it.
You'd be writing thousands of lines of code that do nothing but keep your data in sync with widget states. You'd have many one off bugs where one sub field fails to sync in some scenarios. Your only options is to be more defensive: more events, more full-resync. As a result, the codebase is complicated and the application feels slow/heavy, because updating widget states is costly.
In immediate mode, none of that matters. You don't have a parallel widget tree.
> In immediate mode, none of that matters. You don't have a parallel widget tree.
No, instead, you need to have your entire dataset in memory, and potentially rebuild your whole tree on every update. This causes quite a lot of problems for out-of-core datasets - you end up needing to maintain proxies for things like items in scrollable lists that aren't loaded into memory yet.
This thread (and post) keeps mistaking immediate mode GUIs with declarative UIs. You can have declarative UIs that aren't immediate: the vast majority of them are, and they are idle most of the time.
> What is it that matters for "immediate mode"? Is it that the UI renders everything every frame? No. It's that you build the UI by describing what it should look like everyframe, based only (or mostly) on the data. This is why React won...
I don't think that's why React got so popular. React popularized unidirectional data flow, which is different than immediate mode rendering. This readme file seems to conflate the two of those.
Now that I think of it, couldn't one argue that React itself is a retained mode UI, since it choses which components to re-render and which not to?
Super vibe coded project still uses an LRU cache (github.com/dboslee/lru v0.0.1) as a dependency, this is exactly the things these llms were supposed to solve right, why the hell do people still add these 100 lines dependencies in their projects?
This looks better than the other native Go GUI frameworks. Frameworks like these are a really good target for coding agents: the biggest impediment to building a new UI framework is the sheer amount of meticulous grunt work needed. So, sure, I'm on board with the idea of languages not normally a perfect fit for native UI work getting solid, generated frameworks.
But: what's the real advantage at this point to having frameworks like these? I can get arbitrary SwiftUI interfaces built very quickly, with a lot of attention to macOS (for instance) idiom, and an automatically generated interface between the SwiftUI app and Go. That works pretty great. Why take the UX hit at all?
I think AIs generally "like" the API surface exposed by Shirei. They seem to understand it pretty well and are able to "vibe code" any kind of UI you ask of them using this framework.
To your general point though, yes, there's a lot of uncertainty about what things it makes sense to build in the AI era. I don't have good answers for what would make sense for 10 years down the road, but at this point of time, I think a project like this actually makes perfect sense.
Now that AI can write the code for us, we need better frameworks and libraries, because a lot of what we have now is quite honestly slop.
You could say that you could get AI to maintain 3 separate codebases of the same application UI, each one targeted to a different platform, and the AI will maintain the feature parity and you won't have to worry about it.
For now, AI can take care of coding, but it does not relieve you of having to pay attention to things. If you create 2 or 3 code bases for the different platforms you target, the maintenance burden still falls on you.
But I don't know for how long the situation will remain like this.
Last year I did not think I'd be using AI for coding. I dug up a few tweets I posted last year saying that LLMs have hit a plateau and will stop improving, and that AI coding is a scam being promoted by charlatans.
I understand the core is the layout engine and a component library? Does the rendering somehow benefit from GPU?
I recently had a good experience creating custom UI based on ebitengine — also a cross-platform Go engine. As it is a game engine, it has this built in game drawing loop, GPU-accelerated, with some cross-platform kb/mouse input handling. And this feels like a good platform to build the layout engine and components on top of. Have you ever considered this? Or how does your approach compare to that of ebitengine? Did you try (and do you position) your library to build custom UI for some underpowered computers such as Raspberry Pi?
The needs of desktop and mobile are different enough that it's extremely difficult to build a UI framework for both that doesn't seriously compromise one paradigm or the other.
I would argue it's one of the main reasons why frameworks like Flutter stuggle with widespread adoption on desktop — it was conceived primarily as mobile-oriented, and so on desktop you're stuck with half-baked third party components for essentials such as datagrids and tree views. WinUI with its mobile heritage in UWP suffers similar problems.
GTK + Adwaita tries to straddle the fence and produces a subpar experience on both sides. Desktop data density is terrible due to mobile-minded button sizes and margins (big touch targets, bloated whitespace to make inadvertant touch interactions less frequent) and desktop-oriented widgets like tree views feel out of place on mobile.
they are useless too. 99.99999% of normal people do not open terminal in their entire life nor do not even know what that is. get out and touch grass, friends.
The following are known issues and limitations that we plan to tackle:
Large text blocks will kill responsiveness! Use the LargeText widget.
The widget catalog is aimed at developer tooling, not general consumer polish: no rich text, tree widget, or date picker yet.
There is no robust theming system. Some widgets take an accent color; custom button styles mean implementing your own (the stock Button is a usable reference). Styling can still be verbose at times.
I had a look through the code to see how it manages not to use a pile of C libraries with cgo like other Go GUI libraries.
The answer is platform dependent:
Windows loads the relevant DLLs by hand and calls them. This is a well established technique in Go programs and due to the super stable DLL interface works well.
Linux has an x11 and Wayland backends and these implement (through a library) the wire protocols directly in Go which is nice and will make cross compilation and distribution easy.
macOS does appear to use cgo to access the cocoa libraries. macOS doesn't like statically linked Go programs anyway though as they don't use system name resolution so this isn't a bad compromise, but will mean macOS stuff needs to be built on macOS I think.
I didn't see Android or iOS support.
A nice innovative approach to GUI building. Since the lowest common denominator for the backends is an RGBA buffer, this will bypass all accessibility things the OS provides.
The above gleaned after a few minutes reading the source so may not be 100% accurate.
Theres is another new project which does NOT rely on CGO https://github.com/gogpu/ui Its basically the whole WebGPU implemented in Go. Theres also Gio, but not sure if its still active. CGO is a huge pain…
> Mobile is under consideration (no decision yet). If it is supported, it will be limited to utility-style apps — not games or rich multi-touch experiences.
https://judi.systems/shirei/
No Multiple Windows so even desktop apps will be limited to "utility-style apps".
I compiled and ran the process_monitor example on linux: it works, compiles fast and is about 10mb. Also cross-built for windows and it's 8.4mb. Can't build for macos/arm64
(Under wine the windows exe doesn't render text. weird.)
You need to have some fonts installed. I think for some reason a bare wine setup either has none, or has fonts that Shirei cannot recognize.
examples/process_monitor$ GOOS=darwin GOARCH=arm64 go build
I will admit that I don't like vibecoded things, but perhaps I must stomach that AI will be writing a lot in this brave new era.
However, when the commit history has stuff like
it's very hard. These “change the entire world” commits make for a history that is impractical to follow for a human, and therefore of little interest to me.I think a challenge for a vibe coded library to gain adoption is if there is a lack of human time investment in its creation, how do we know there will be investment in its maintenance?
Most open source projects are abandoned because the author never finds enough time to work on it or can't muster the focus and attention needed.
It is not true that there is "lack" of human investment in the creation of this. If anything, I spent the last two weeks glued to the screen most of the time, to a degree I have never experienced before, building out all the different areas that lead to this release.
I will not mention the monetary investment because it's not the type that matters here.
Attention, which is arguably the most scarce form of investment, was invested in ample amounts.
> I spent the last two weeks glued to the screen most of the time, to a degree I have never experienced before,
With respect, a single period of manic hyperfocus is not what makes a project, it is being in it for the long-haul and putting in the hours here and there even when it isn't fun any more and you have many other things you might rather be doing.
To be clear I don't want to detract from what you've made here. Agentic engineering done right does still require a lot time and attention IMO. But you've still got a challenge on your hands to demonstrate to potential adopters that you're in it for the long haul.
This is a publish-only mirror repo.
The commit history is the publish history, not the work history.
but if you other ppl to read, use or contribute, they will probably need the "work history" more than a "publish history". because nobody is going to read those big ass commits, specially when unsure about it worthiness.
Shirei is open source, not open contribution.
Where's the actual repo then?
that doesn't make much sense. why go out of your way to publish a selected history instead of the whole work?
I have several projects in the same git repository which also forms a workspace in Go. Many of the projects in that repo are not public.
ok, so because of that you manually export changes you want to publish to a new repo? i'd start separating the repos. personally i would be uncomfortable using a repo that the developer does not use for their own development. if you ever want to accept contributions this might become an issue. but that's just my feeling.
That's irrelevant though? I'm not asking anyone to contribute or form a community with me as the leader.
I'm putting out something in the hopes of it being useful for others.
I'm assuming you're the author, why not publish everything?
Source is the new binary. Specs and requirements are where the engineering is happening now. Of course, this was anticipated in 1971 with the creation of the very successful PRIDE methodology, but it's taken a few decades, and a complete devaluation of their profession due to AI, for programmers to catch up.
This PRIDE? https://cio-wiki.org/wiki/PRIDE_Methodology
PRIDE is the registered trademark of M. Bryce & Associates (MBA)
https://www.modernanalyst.com/Resources/News/tabid/177/ID/50...
and https://web.archive.org/web/20240422044352/http://www.phmain...
https://patch.com/florida/palmharbor/50-years-pride has the meaning "an acronym for "PRofitable Information by DEsign - through phased planning and control."
This is the PRIDE I meant. It's hard to tell if MBA (more recently, M&JB Investment Company) is still a going concern. Milt Bryce's son Tim was keeping the lights on but he died of cancer at the end of 2023.
> Experience has shown us that an immediate mode API is the only sane way to program GUI applications
I wonder how long till they pivot away from this belief. I feel like everyone in UI goes through this phase as some point, but in the end it doesn't scale to truly complicated UI
Yeah. Based on my personal experience I think some kind of hybrid of old-school imperative retained and declarative retained, both with granular reactivity is probably the correct balance for "serious" high-utility desktop applications. Declarative approaches are great for smaller components but become a nightmare for anything much more complex than a relatively simple mobile app while imperative requires a lot of extra legwork at the component level, and as I understand (which may be incorrect) immediate mode makes certain types of optimization more difficult.
In my experience, it's the opposite. Immediate mode GUI, or at least a functional and declarative approach, is the only way I've seen it scale well. It's more modular and scale-independent. On the other hand, retained mode, or imperative/OOP approach to state management, becomes complicated and monstrous quickly; it's the dominant style and can be made to work OK, but typically hellish to maintain or develop beyond a certain scale.
Admittedly I'm simplifying too much and conflating paradigms. My preference is something like "functional core, imperative shell" or maybe "immediate-mode core, retained-mode shell" if that makes sense.
Is Tracy complicated enough? Because it's imgui.
https://github.com/wolfpld/tracy
That looks quite simple. Think about something like this, a commercial SystemVerilog simulator (this only shows a fraction of the UI).
https://blog.reds.ch/wp-content/uploads/2018/09/questa13.png
Or something like Visual Studio.
Obviously most GUIs are not nearly that complex so immediate mode can get you quite far. Its biggest limitation is that it makes it hard to do some layouts. Your GUI layout becomes dictated by your data dependencies which is quite awkward.
Absolutely zero difficulty redoing this in a react style renderer. The only complexity is being careful with your data dependencies so as to not needlessly rerender.
Each pane is easily isolated, can share data with a view model scoped properly, etc. Writing it in an imperative toolkit is a "oops I forgot to update my data here" kind of hell. Data binding makes it slightly less worse
> Absolutely zero difficulty redoing this in a react style renderer
I would not classify a react style renderer as "immediate mode". It has aesthetic similarities with IM GUIs, but ultimately there is a fully retained tree under the hood, that gets diffed/mutated on every update
It's the only thing that can scale to complicated UI
What does "scale"even mean in UI context? 10 or 100 controls in app makes difference how exactly? Retained apps redraw when needed, they are idle most of the times. How redrawing every frame helps to scale?
Does "immediate mode" mean the UI refreshes at a constant 60fps rate like a video game, redrawing everything all the time?
No.
It means that you build the UI by describing what it should look like now, based on the data / state you own, without referencing any existing "widget" object or trying to manipulate it.
Scale is not about the number of buttons, but the structure of the data.
You have a list of objects, within each objects you have several fields, some of them lists, some of them maps. within some of those sub-items you have other lists and maps, nested arbitrarily.
This would be hell to manage for a retained mode UI. You have to mirror the application data into a widget tree and keep all the elements in sync, all the way down to the arbitrary depths of it.
You'd be writing thousands of lines of code that do nothing but keep your data in sync with widget states. You'd have many one off bugs where one sub field fails to sync in some scenarios. Your only options is to be more defensive: more events, more full-resync. As a result, the codebase is complicated and the application feels slow/heavy, because updating widget states is costly.
In immediate mode, none of that matters. You don't have a parallel widget tree.
> In immediate mode, none of that matters. You don't have a parallel widget tree.
No, instead, you need to have your entire dataset in memory, and potentially rebuild your whole tree on every update. This causes quite a lot of problems for out-of-core datasets - you end up needing to maintain proxies for things like items in scrollable lists that aren't loaded into memory yet.
> you need to have your entire dataset in memory
Are we talking about displaying tables with billions of entries? I'd like you to show me how retained mode gui scales well with that use case.
This thread (and post) keeps mistaking immediate mode GUIs with declarative UIs. You can have declarative UIs that aren't immediate: the vast majority of them are, and they are idle most of the time.
> What is it that matters for "immediate mode"? Is it that the UI renders everything every frame? No. It's that you build the UI by describing what it should look like everyframe, based only (or mostly) on the data. This is why React won...
I don't think that's why React got so popular. React popularized unidirectional data flow, which is different than immediate mode rendering. This readme file seems to conflate the two of those.
Now that I think of it, couldn't one argue that React itself is a retained mode UI, since it choses which components to re-render and which not to?
Super vibe coded project still uses an LRU cache (github.com/dboslee/lru v0.0.1) as a dependency, this is exactly the things these llms were supposed to solve right, why the hell do people still add these 100 lines dependencies in their projects?
Rebase (squash), I assume
This looks better than the other native Go GUI frameworks. Frameworks like these are a really good target for coding agents: the biggest impediment to building a new UI framework is the sheer amount of meticulous grunt work needed. So, sure, I'm on board with the idea of languages not normally a perfect fit for native UI work getting solid, generated frameworks.
But: what's the real advantage at this point to having frameworks like these? I can get arbitrary SwiftUI interfaces built very quickly, with a lot of attention to macOS (for instance) idiom, and an automatically generated interface between the SwiftUI app and Go. That works pretty great. Why take the UX hit at all?
I think AIs generally "like" the API surface exposed by Shirei. They seem to understand it pretty well and are able to "vibe code" any kind of UI you ask of them using this framework.
To your general point though, yes, there's a lot of uncertainty about what things it makes sense to build in the AI era. I don't have good answers for what would make sense for 10 years down the road, but at this point of time, I think a project like this actually makes perfect sense.
Now that AI can write the code for us, we need better frameworks and libraries, because a lot of what we have now is quite honestly slop.
You could say that you could get AI to maintain 3 separate codebases of the same application UI, each one targeted to a different platform, and the AI will maintain the feature parity and you won't have to worry about it.
For now, AI can take care of coding, but it does not relieve you of having to pay attention to things. If you create 2 or 3 code bases for the different platforms you target, the maintenance burden still falls on you.
But I don't know for how long the situation will remain like this.
Last year I did not think I'd be using AI for coding. I dug up a few tweets I posted last year saying that LLMs have hit a plateau and will stop improving, and that AI coding is a scam being promoted by charlatans.
I understand the core is the layout engine and a component library? Does the rendering somehow benefit from GPU?
I recently had a good experience creating custom UI based on ebitengine — also a cross-platform Go engine. As it is a game engine, it has this built in game drawing loop, GPU-accelerated, with some cross-platform kb/mouse input handling. And this feels like a good platform to build the layout engine and components on top of. Have you ever considered this? Or how does your approach compare to that of ebitengine? Did you try (and do you position) your library to build custom UI for some underpowered computers such as Raspberry Pi?
Shirei does not use GPU for rendering. It's fully software rendered.
> ebitengine
I have considered using it as a backend, but the blocker for me was how it handles resizing: the window content will stretch while it's being resized.
How it compares to Fyne?
Wails is another cross-platform GUI framework in Go: https://wails.io/
Wails is more like Tauri (Rust) - https://tauri.app/ - both use a WebView to render the frontend.
What a waste of tokens.
true to that. looks like AI took shortcuts and avoided real work and real hard problems. might as well just flip a random bits for cpu hours.
cross-platform is overstatement.
can I run it on Android? iOS?
no? then 99.999999% of real world users cannot access it. and if it is desktop oly, what is the point? it is no better than web.
The needs of desktop and mobile are different enough that it's extremely difficult to build a UI framework for both that doesn't seriously compromise one paradigm or the other.
I would argue it's one of the main reasons why frameworks like Flutter stuggle with widespread adoption on desktop — it was conceived primarily as mobile-oriented, and so on desktop you're stuck with half-baked third party components for essentials such as datagrids and tree views. WinUI with its mobile heritage in UWP suffers similar problems.
GTK + Adwaita tries to straddle the fence and produces a subpar experience on both sides. Desktop data density is terrible due to mobile-minded button sizes and margins (big touch targets, bloated whitespace to make inadvertant touch interactions less frequent) and desktop-oriented widgets like tree views feel out of place on mobile.
so it is not cross-platform? or all available platforms are... desktop.
serious vibes of shortcuts and avoiding real hard work that brings real value.
It is still meaningfully cross-platform to support all 3 major desktop platforms rather than just one.
If desktop only is so useless, why are so many people create TUI apps?
they are useless too. 99.99999% of normal people do not open terminal in their entire life nor do not even know what that is. get out and touch grass, friends.
Web is fantastic.
> Immediate mode API in the true sense: you never need to maintain UI widgets or sync your data with widget state.
https://judi.systems/shirei/
Known Issues & Limitations
The following are known issues and limitations that we plan to tackle:
Are these statements compatible?