I highly recommend everyone to give htmx or datastar a try, especially if your main experience is react or nextjs.
We recently rewrote a half-million LOC codebase from react to datastar, with a detour through htmx first, and the results are staggering.
First page load is 20KB down from a 750KB js bundle. 1 network request vs 40+. Total load time 0.1 seconds down from 2 seconds of spinners. Page refresh is so fast, the browser doesn’t even flash - there’s no massive js bundle to parse, so it can start rendering instantly.
But the most staggering result is that in-document navigation over SSE achieves up to 5000:1 compression ratio for network requests within the Brotli compression window. Yes, 5000:1, you read that right. When changing server state we rerender the entire page, like an immediate-mode game engine, and push it to the client over an SSE “fat morph”, in about 100 bytes.
It’s so fast and so much more stable, our users are literally having fun. With a medical device.
And this ends up ~50% less code to maintain for a VASTLY superior UIX.
You are really missing out if you don’t learn how this works. It can give you a significant competitive edge if you care enough to try.
Edit: Big BIG thanks to bigskysoftware and the datastar team for making the web a sane place again. You really deserve all the props.
My experience was similar. I'd wound down a startup and as a solo developer could no longer afford to do react + a backend. So I tried HTMX, then turbo (which introduced me to morph and push based re-fetch). Finally moved to datastar because I could drop alpine and it has best in class morph, signals and SSE support. Immediate mode over the network is amazing DX.
Makes multiplayer out of the box over large datasets trivial. See these demos if you don't believe me [1][2] (open them in multiple tabs and have fun)
Our UI is highly interactive: think 3d visualizations with toggle-able overlays, various display settings, searchable lists, drawing tools etc.
We made a lot of progress with htmx, but eventually hit a complexity wall where we were spending more time fighting the tool than solving the business need. Note, I'm not blaming htmx here, the documentation is clear about its strengths and limitations, and the hypermedia approach is still fundamentally right. But even with alpinejs in the mix, we just weren't able to achieve the smooth developer experience we were looking for.
So we tried datastar, and it basically addressed all the friction points we had:
- No more DOM hunting to figure out why the hx-get ended up in the wrong spot or why OOB swap is glitching out. D* morph just works.
- No need for hidden form <input> to submit client state, D* signals are cleaner and easier to use.
- No more micro-routes for targeted element updates to avoid hx-swap butchering page state. One route per page, just render the whole thing and morph takes care of the rest. Immediate-mode rendering and fat morphs removed ~50% of our routing table.
- SSE + fat morph is a superpower. This is view = f(state) done right.
We've had very few issues, D* just works and gets out your way. The team has really created something quite exceptional.
The great news is htmx4 is adopting many of these ideas, so I recommend trying both and seeing what fits you best.
It's nice to see HTMX4 is adopting the core strengths of Datastar. That sort of cross pollination of ideas is a huge boon to the wider hypermedia community in general.
HTMX v4 doesn't actually address what's makes Datastar work at it's core. It at best it copies some of the broad strokes of the API but not in a robust it efficient spec compliant way
A focus on morph by default means you don't have an explosion of endpoints and can write programs like the 300 redirect days. [1]
Once you want realtime/collaborative app SSE + morph lets you do that easily. Add streaming compression and this approach is suddenly crazy bandwidth efficient.
For people coming from react this gives you v = f(state) over the wire.
Lastly signals handle the edge cases where an out of bound morph would nuke your ephemeral client side state (mostly text input).
You can do this in htmx but you are fighting the design the whole way. Turbo is a bit better but doesn't have signals.
I’m getting about 800:1 compression with Brotli. I wonder if need to tune anything? Although tbh I have no performance issues at all.
It took me a minute to change my mindset to Datastar’s way of doing things. But it’s been amazing. Firstly to keep 99% of the state and logic on the back end is just so nice - no more having to deal with two sources of state at the same time (front end and back end). Secondly the ability to regenerate the entire view from scratch anytime something changes and then have Datastar efficiently morph just the changes is just such a simple mental model. Every fiber in me was like “no way that can work, the cpu, bandwidth, latency etc” - but it is buttery smooth.
So compression ratios with streaming compression (brotli/zstd) come down to the regularity of the data, the size of the change, and the size of your window.
Basically, if your html fat morph frames are 65kb uncompressed and you make a small change like checkbox it will send 13bytes so that's 5000:1. Caveat is as long as your compression context window is big enough.
You either buy a datastar license to get their debug tools, or you spend a couple of hours to build your own. I'm sure you can find something on GitHub too.
One of the cool things is how extensible d* is. Everything is a plugin, and you are free to extend it to suite your needs. We've invested a few days building a few specific data-* attributes for our application and it's super convenient.
Why are you so wedded to JSON responses? HTMX is meant to be used to render html; my big HTMX app basically never(1) renders JSON, so much so that I have a standing instruction to coding agents that if they're rendering JSON they're probably doing something wrong. If you're attached to JSON because you're working with external API s you don't control, a good place to perform those kinds of actions is in your backend, which can then return html to your frontend
(1) the exception is I think in one place concerned a flow for uploading images, where client side requests a signed url, uploads to that url, then registers completion. Getting the signed url back is JSON meant for the uploader
But now I want to add a JSON API to the program: I could reuse the same routes (/create, etc), but require an application+json header; or add a full /api/ component essentially duplicating the logic from the main app.
Fat services, thin routes. Try moving your business logic to a library and consume it from both your UI and API routes. That minimizes duplication and decouples your public API resolvers (with backwards compatibility requirements etc) from your private UI resolvers (which you want to evolve rapidly to optimize your UIX).
That's what we did and it works pretty well in practice.
As a result for POST requests you have to have htmx routes that accept urlencode and API routes that accept JSON, and JSON is far superior. But for outputs I actually think separating your UI and API is helpful. The code reuse is not worth the entanglement of concerns.
Htmx is not meant for these use cases. We’re building great interactive webapps with htmx using the web platform’s built-in capabilities. If you need exact scroll position and text highlighting you can get that with some other framework. Htmx works well with things that are in its own niche.
Because I would need to see a UI paradigm prove it can handle performance before considering using it for UI.
react-three-fiber solves a lot of problems. React state is also perfectly in line with game loop architecture, r3f unifying the Three.js/React render loop is incredibly good for game dev.
But even for SaaS or whatever, I need to know that it can handle complex states of the UI - HTMX can't.
They are overselling it. It's not really a competitor to React, it's more like a competitor to Jade templating or HAML which nobody already uses anymore.
If I have to build something that is basically logic on backend, forms, tables, some static pages, some interactivity here and there(btw is what most of the sites and web apps out there are) I don't need 3d or game engine capabilities ...
And no way react is going to be faster for the above use case.
You miss the point of tools like htmx/datastar/turbo/unpoly/alpine, they are not a react competitor in the rich and complex web apps space. They fill a gap between the native browser capabilities and tools like react/vue/svelte/ecc.
HTMX makes it hard to build even basic things, because it swaps entire HTML forcing a full repaint. That's the foundation of how it works.
Examples include user highlighting text - swap wipes that out. Or user mid scroll through a menu, the scroll is reset to the top.
It's not enough in HTMX to just break it down into smaller components, the scrolling part is native to the browser, so is text selection.
Some people forget or don't know how much an SPA library like React is doing to prepare the SPA before you get into any of the organizational framework usage. Beyond that, clever engineering went into React DOM reconciliation in particular that I rarely see challenged in other libraries.
On game dev: I think it's good for a showcase because it shows the limits of what it could do at 60+ fps in a performance intensive environment. But yeah, even a complex SaaS demo would do.
> They fill a gap between the native browser capabilities and tools like react/vue/svelte/ecc.
Have not heard that yet, since a big part of HTMX is manipulating the DOM and binding events. I'm pretty sure this is not correct
I'm using Hono + Tanstack Router as a Next.js replacement. Best software related decision I ever made, it feels like I stopped fighting with the stack.
Bravo! I never would have thought this would be so simple. As a direct consequence of this exercise: I now know something about what, precisely, htmx does.
Huh, so every interactive event has a network request in the loop, with perceptible latency? And you still end up obligating client-side JS for the plumbing? I'd been meaning to look into HTMX but based on the content of this post it seems like a worst-of-all-worlds technical solution. Is it at least pleasant to work with?
No, every interactive event doesn't need to have a network request in the loop. And you only need a tiny bit of client-side JS for a most of the plumbing.
It is quite pleasant to work with, because you don't have to worry about a constant stream of vulnerabilities or supply chain attacks. You just vendornthe script in your repo, add it as a <script> tag on the page, and you're done.
As is well known, security is not a binary, it's a spectrum. It's why you add security in layers, and why it's generally a trade-off between risk and usability.
Your mileage may vary, but I sleep better with my own projects not having any node_modules in their build tree.
No, it's not specifically npm. All package managers enable remote code execution by design - you tell it to download code and then you run that code, likely without sandboxing, by using it. Most people are not inspecting the code before they download and run it. You will still see compromised packages making their way through npm even though they disabled install scripts.
That's why many times HTMX is used together with alpinejs [1].
For areas that need server side, if you use say react and return a JSON you have to deserialize and do the render. It might be much more efficient to just replace.
Having said that I use HTMX /alpinejs for small projects, and still trust react for bigger more interactive ones.
It's for building extremely backend-brained frontend stuff, like a sort of anti-NextJS. Also like NextJS, this sort of domain-bending has some unfortunate operational consequences.
Datastar has a lean and mean signals implementation that handles all the front-end reactive stuff - so you don’t have to go to the server for interactivity.
Is there something known wrong with github rn (aside from everything, I guess)? It claims there are 422 contributors but when you click in there are only three.
GitHub's contributor graph is based on authorship of Git commits, and doesn't track whether a given GitHub account ever actually interacted with a given GitHub repo. This Git repo is a fork of the htmx Git repo (though it's not marked as such on GitHub), so all contributors to htmx as of the fork date are also contributors here.
Ahhh makes sense if it's a fork. Odd that it's not marked as one. Sort of an odd thing that you can get associated with a repo you don't even know exists.
We recently rewrote a half-million LOC codebase from react to datastar, with a detour through htmx first, and the results are staggering.
First page load is 20KB down from a 750KB js bundle. 1 network request vs 40+. Total load time 0.1 seconds down from 2 seconds of spinners. Page refresh is so fast, the browser doesn’t even flash - there’s no massive js bundle to parse, so it can start rendering instantly.
But the most staggering result is that in-document navigation over SSE achieves up to 5000:1 compression ratio for network requests within the Brotli compression window. Yes, 5000:1, you read that right. When changing server state we rerender the entire page, like an immediate-mode game engine, and push it to the client over an SSE “fat morph”, in about 100 bytes.
It’s so fast and so much more stable, our users are literally having fun. With a medical device.
And this ends up ~50% less code to maintain for a VASTLY superior UIX.
You are really missing out if you don’t learn how this works. It can give you a significant competitive edge if you care enough to try.
Edit: Big BIG thanks to bigskysoftware and the datastar team for making the web a sane place again. You really deserve all the props.