XMPP vs. Matrix
I want to know everybody’s opinion on what’s better to use today, XMPP, or the Matrix protocol? Maybe also what’s easier to convert your friends and family to.
XMPP was invented at a time, where communicating online meant sending a message from one device to another.
However, the modern expectations for messaging apps are much more than that. Sending media, using multiple devices, deleting messages, editing messages, read receipts, notifications when typing, group chats, threads, and even managing communities are all things a modern messenger app should be able to do. The fundamental operating principle has shifted from mere message passing to synchronising a common state between all participants. If you think about it, nowadays, you're not chatting anymore. You're essentially collaboratively editing a shared chat history file, where the most common action is to add a message; usually at the bottom.
This is what Matrix is at its core. It's a protocol to synchronise state, and that's part of why Matrix is so complex and hard to administrate. I personally think its the better base for the future of communication than XMPP, and I havent even mentioned encryption yet.
Moving on to the practical part: Running a Matrix Synapse server is quite a commitment, but if all you want is talk to friends and family, then there are simpler options. Conduit and Dendrite are a bit easier to set up, and of course there are plenty of public homeservers you could sign up with.
If you do commit to running Synapse however, you have the option to install bridges to almost any other messaging service. This way, your friends and family can keep using what they're used to (WhatsApp, Telegram, Discord, Facebook, ...), and you just use one single Matrix client to talk to all of them.
That's what I do.
> XMPP was invented at a time, where communicating online meant sending a message from one device to another. However, the modern expectations for messaging apps are much more than that. Sending media, using multiple devices, deleting messages, editing messages, read receipts, notifications when typing, group chats, threads, and even managing communities are all things a modern messenger app should be able to do.
XMPP provides all of these features and manages to keep up with commercial products really well. Everything Slack or Discord offer is there in the XMPP protocol. And if it wasn't, it could be relatively easily added, thanks to it being extensible.
However, navigating the protocol and software supporting it requires a little bit of know-how. If the OP is interested in building a product incorporating instant messaging and the satellite features, I'd suggest partnering up with somebody with this know-how. Scalable servers would be MongooseIM or ejabberd, polished clients are Conversations or Movim.
If it's a question about which protocol to use for a homeserver, then maybe something focused on ease of setup would work best, like Prosody.
> The fundamental operating principle has shifted from mere message passing to synchronising a common state between all participants.
So it should all be based on blockchain, shouldn't it? ;)
Commitment in what way? I found it fairly easily to set it up with a domain of my own.
This ansible playbook helps a lot but it's still periodic annoying maintenance and it's still way more resource intensive than it should be: https://github.com/spantaleev/matrix-docker-ansible-deploy
EDIT: I run it on a hetzner dedi, I have run it on Apple fusion drives in the past and consumer SSD's and it was a terrible time. Need high IOPS so things don't grind to a halt on intensive db operations (like joining HQ channel lol)
on top of that not a month goes by where something doesn't break.
a security conscious friend who joined matrix because of me (well, he was interested before, but i finally gave him a reason) just deleted his matrix clients in disgust after some messages randomly could not be decrypted.
i suspect that there is still a problem on the server i am using, therefore more potential work for its already overworked admin.
Your comment makes it seem like they don't.
XMPP is traditional IM. I use it exclusively from Conversations client on Android / GrapheneOS and it is really instant, supports presence (knowing whether contact is actually online), rings for audio and video calls, gives feedback about whether user read the message etc. I use it mostly for 1:1 conversations, it has all but replaced SMS and phonecalls.
I consider matrix more like 'fast forum'. Perhaps things changed but last I checked (from element on Android / GrapheneOS) there was no presence so I have no idea whether contact will get my message immediately or not. No confirmation that contact read the message. Audio and video calls not working, not even ringing when phone is locked. Quite laggy in message delivery.
So, after some years of using both, XMPP is best for replacing one-to-one SMS, video and audio calls, while I enjoy hanging in public matrix rooms, treating them like '(not really) instant forums'.
- I'm now sure that I will never convert anyone to Matrix or host a Matrix server again.
- Matrix big issues for me are that it's not really decentralized and this design impacts the server admin, the performances and the user's privacy.
- I'm still not happy with XMPP clients for Linux, most are missing Omemo or automatic turn/stun discovery and Dino is in in alpha and bugged.
- I'm sur that XMPP will have a community in 20 years, not sure about Matrix.
Thanks!
> Matrix isn't really decentralized
I think most of the concerns stem from the majority of the development in the ecosystem being driven by a single company. This manifests in a few ways:
matrix.org is by far the biggest server. Element is by far the most popular client. Both were (until recently) maintained by the same company, and the client selects matrix.org as the default server. This was done to make onboarding easier, but it centralised power into one organisation.
Theoretically, since Element does most of the development, if they want to push something into the protocol, the whole ecosystem really has to follow.
Because most people are on matrix.org, and the protocol works by replicating state between involved servers, matrix.org gets a lot of metadata even from conversations that are not hosted on their server (as long as one person in the chat is on matrix.org).
There are some flags which are enabled by default which automatically "phone home" to matrix.org even when you are running your own server. They are easy to turn off, but an annoyance for people who want to run completely separated instances.
The standard way to join a chat is via a "matrix.to" link. This leaks some information to matrix.org even if the server is completely separate.
By contrast, XMPP has no dominant player who drives most of the development, and no dominant server instance. I imagine this is because the protocol has been around for much longer (so there is more time for things to settle), and there is a very clear separation between the standards body and the software developers. There are popular clients that have some influence over the ecosystem, but nowhere near the extent as in Matrix.
Now, XMPP has its problems too, but few of these stem from having too much centralisation.
> which XMPP client you're using on Linux
I personally use Dino (https://dino.im/). It offers a simple, clean experience, and looks really nice.
There are also more fully-featured clients, like Gajim (https://gajim.org/)
* the webserver hosting matrix.to can see your browser’s IP and user agent as it connects to it. (this could be improved by connecting to matrix.to via an overlay network of some kind, or simply using matrix URIs instead)
* it cannot see the room/user being linked to, as this is deliberately hidden in a url fragment from the server
* it cannot see details about the room/user being linked to (eg the icon or room name); it optionally lets the client chose to grab this from a given homeserver, but the service itself never sees this data.
In other words, it was written with privacy in mind.
That way for people who don't care everything just works and for people who do they get a rich experience with E2EE. Unlike eg iOS/imessage I don't force them to use a different OS for this and I can talk to everyone from my laptop. Everything is 100% self hostable.
I don't know why you would use Matrix. Self hosting is awful, the clients are a huge mess, and it was bootstraped by the Israel based company that runs US telephone surveillance (it's kind of insane that's done in another country.)
That argument, again.
It's been some 7 years since Amdocs withdrew all funding, forcing the devs to create their own independent thing. Since then, as far as I'm aware, the protocol has been developed in the open, with lots of external contributors. There's plenty of implementations, both client and server. How are its roots relevant? Are you going to say the internet is dubious because it was initially researched on by the US military?
Anyway, your other points are relevant, though it would be nice to have some more details.
Regarding features, matrix is promising and definitly innovative, but espacially the mobile apps don't have the same level of usability like WhatsApp or novel features like Telegram. Techsavvy friends can definitly use it, but you don't want to become a managed service provider for your broader family.
I use this ansible playbook to provision my server and related services (monitoring, bridges, ...) [0].
The bridges espacially make it fun to play around with.
[0] https://github.com/spantaleev/matrix-docker-ansible-deploy
It's quite critical of some of the code quality of common implementations as well as the fracturing across different clients.
As for Matrix, probably element is the main client you want to use. I use Nheko on Linux.
He keeps criticizing one client while trying to pass it off as criticism of the specification (said client doesn't even implement the latest version of the specification, either), complains about the protocol changing too much and the protocol changing too little, and to top it off seems to promote Signal as the better alternative, completely missing the elephant in the room -- Signal being centralized with its decisions even more whimsical at the hands of one group only. Curiously, he does that right after criticizing omemo's choice to follow Signal's choices as lacking justification.
There are two major problems, the implementations and the fact the spec isn't specific. I'm not particularly bothered by the imaginary, the dude is a furry what do you expect? Some furry bloggers do have pictures throughout their blog posts to split things up and lighten things.
As for the reply from the spec author, I had a negative opinion of that. To me it looks like they just tried to copy signal's encryption so that people could claim that XMPP has E2EE, without any real design or thought of the implementation.
One of the other things I hate about XMPP (and you mentioned it above with the jingle debacle), but there are other cases where there's multiple XEPs for the same thing. Some of them are widely used despite being marked as "experimental". The documents are not cohesive and is a mess which is probably why the implementations are also bad (unlike the Matrix spec). Encryption is just another one of those things, just like file transfer.
If I was deciding on a new project to develop a client for XMPP would be the least attractive project for me to work on. Put it that way. Without enthusiastic developers that think this thing sounds cool there simply won't be any good software.
The other thing also not mentioned is that the OMEMO encryption only applies to text messages and not all the other things, ie VOIP, status changes etc.
There is a fair bit of metadata on the server side
https://web.archive.org/web/20211215132539/https://infosec-h...
and attacks like this are just downright scary
https://notes.valdikss.org.ru/jabber.ru-mitm/
I used to use XMPP but haven't in about a decade. Nobody I know uses it either. (Even the couple of evangelists I knew moved to Matrix long ago)
Lighten things? Are you saying that putting images of cartoon characters puking at the logos of your product lightens things and provokes healthy discussion? Don't try to make this into thinking I'm criticizing furriness -- it's definitely not about that.
> There is a fair bit of metadata on the server side
This is much less critical when the protocol is federated. Even E2EE itself may become second priority rather than first in such an environment.
> Without enthusiastic developers that think this thing sounds cool there simply won't be any good software.
This is a ridiculous statement.
False. Federated protocols typically have a lot more, as there is routing data between servers. The same goes for things like Matrix.
The only thing that Matrix servers can see however realistically is the room id, and other participants. Unlike XMPP where they can copy your roster. In the case of ejabberd debug logs can include your password lol (or at least could in November 2021). I doubt that's changed.
This isn't always the case, not every user is a server admin. Also for group chats other server admins will certainly know what data your requesting.
TLDR is that XMPP is not a private protocol, and that was never its intention. It was actually very centralized in it's original usage and the designers clearly envisaged similar usage to email ie $companyA.com employees talking to $companyB.com not some sort of "private messenger" competitor to Matrix or Signal which were meant to be used by the wider general population.
> Leaking metadata to a centralized service is simply inevitable by pure physics
Not necessarily true https://signal.org/blog/sealed-sender/
Signal has a lot less metadata than XMPP.
> Not necessarily true https://signal.org/blog/sealed-sender/
Yeah, no. Timing alone is enough to clearly distinguish a sender, and any attempts like these is pure astronaut engineering to obfuscate something that is trivial to recover by a million side channels.
There is just No. Way. to eliminate the problem of metadata leaks to a centralized server, other than making it practically distributed/P2P. The more time it takes for people to realize this, the worse society becomes.
Yes it does, because the threat model there is not the server itself. You trust that as it's your employer's server and you're using it for work related purposes.
> Timing alone is enough to clearly distinguish a sender
It's a lot harder to pull off than doing an MiTM attack with STARTTLS and then just observing the person's roster being sent to them when they connect lol.
Yes, that is the point. Or it is my home server sitting at my router serving my family. I have been arguing that if I can trust the server then the discussion about E2EE becomes less relevant.
> It's a lot harder to pull off than doing an MiTM attack with STARTTLS and then just observing the person's roster being sent to them when they connect lol.
And here you make the dangerous mistake that I most hate. It's not only trivially easy for the server operator to leak metadata, it's actually HARD to avoid it.
That works in your situation, but not for most people who don't want to have to maintain a moving part. Also hosting in general on a residential connection can depend on the provider and even availability of good internet.
For a lot of people it wouldn't even be possible if they wanted to and that's not getting into the technical investment they would have to make in knowing how to do such a thing in the first place.
> And here you make the dangerous mistake that I most hate. It's not only trivially easy for the server operator to leak metadata, it's actually HARD to avoid it.
It's a lot easier than simply not having that data server side in the first place like Signal for example.
Frankly, that works for the majority of XMPP users -- let's not forget about that. But yes, this is a problem today for many consumers. However, I see this as a much better direction for the Internet as a whole to take -- see efforts such as ownCloud and the like -- , rather than just conceding defeat and accepting a centralized Internet, or worse: the extremely dangerous idea that centralization increases security. Which should only be seen as the non-sense it is.
> It's a lot easier than simply not having that data server side in the first place like Signal for example.
You still do not get the point here at all. This is not about storage. This is not about the server implementation. As long as there is a communication at all between me and the server and then my contacts and the server, it is irrelevant if you are storing the roster or not. ANYONE (*anyone with access to that server) can easily pick it up. In fact, it is HARD not to leak it up, even by accident. Barring a P2P Freenet-like thing, this is _unavoidable_. (and even with P2P/Freenet/Onion/whatever it's not trivial) And due to the nature of IMs, metadata is practically as big an issue than protecting the payload itself.
I think you can still have decentralization without having to be a sysop. That may be server operators that run small servers etc.
Decentralization doesn't inherently mean privacy though and it's important to remember that.
> You still do not get the point here at all. This is not about storage
Pretty sure signal servers can't intercept messages or decrypt them. In regard to the client you cannot send unencrypted messages. We do know that lawful interception warrants sent to Signal are not particularly effective only registration time and last connection time is available. https://signal.org/bigbrother/
Please, I'm really trying to put the emphasis on metadata, and definitely Signal servers can intercept that at will. Even here on HN there was recently an article of how "lawful interception warrants" were targeting Apple's and Google's _push notification servers_ (those _by design_ really only have access to metadata -- that should speak about how important metadata is for law enforcement). The Signal server has access to significantly more metadata about you and your communications than your typical push server. Again, there's just no contest here: simply avoiding the issue of using their server in the first place is much better than anything they claim* to do (or even better than anything they could _physically_ do).
* I absolutely detest these type of "our privacy is great, believe us!" pages. Time and time again it shown that they are absolutely useless. Where did Apple report that their push notification servers were tapped? The fact that a server collects enough data about you that it is a tempting target for authorities should be cause for concern, not relief.
This never effected signal as no data that is sensitive was ever sent via that. Again though warrants against Signal are not new, and for years law enforcement has been getting the same answers to their subpoenas. If it was "possible" as you say and that "they do", pretty sure the ACLU would not be making submissions to a court that "this is all that is available <registration date> & <connection date>" especially as the source code is open, it would be easy to verify if that was in fact true.
> I absolutely detest these type of "our privacy is great, believe us!" pages
The page literally has the legal request and reply, these would be able to be checked against court records. Remember being in contempt of a court order is actually punishable, so if there was more I'm sure Signal would have been fined by now lol.
How are you still missing the point that this is _about metadata_? No one is sending "sensitive data" through a push notification in the first place. You have to go out of your way to do that. My point was to show you that the authorities _are_ after metadata.
[1]: https://www.ndss-symposium.org/ndss-paper/improving-signals-...
[2]: https://arxiv.org/abs/2305.09799
[3]: https://github.com/simplex-chat/simplex-chat/issues/4620#iss...
https://github.com/simplex-chat/simplex-chat?tab=readme-ov-f...
but the first 2 points were already implemented as shown in their Roadmap[1], so they are going to probably remove them
[1]: https://github.com/simplex-chat/simplex-chat?tab=readme-ov-f...
The image in question does not contain puking. That's a tongue, not vomit.
https://bunnypa.ws/sticker/CAACAgEAAxUAAWByYjqChwgli1QKv81yw...
Here's another from the same artist: https://bunnypa.ws/sticker/CAACAgEAAxUAAV91eRTulCrBXcKzxCrPY...
It contains a disgusted reaction. The logos in question are visually "muddying the waters", too.
https://notes.valdikss.org.ru/jabber.ru-mitm/
That attack strikes me as generic. As in not anything to do with XMPP specifically.
The fact that STARTTLS is even possible with that protocol is bad.
> https://notes.valdikss.org.ru/jabber.ru-mitm/
With an up to date Conversations on a modern server we have a pretty good chance to detect or prevent that style of attack due to a mechanism called SASL Channel Binding.
Note that the author of the linked article complains about getting the same sort of article in reply. That is also common...
Can you point out where this happens? I didn't come away with this impression at all.
Signal, Matrix, Telegram, XMPP; Use whatever you want. But there is a lot of FUD if not outright lies in that blog post. The author looked at Conversations for all but five minutes, desperately trying to dig up some dirt.
For example...
* The auth tag truncation was 'silently' introduced in the spec. It wasn’t. The author retracted that but only barely
* ominously pointing out that Conversations has a SASL implementation (In fact Conversations can use that to detect some MITM attacks; which is pretty cool)
* ominously pointing out that Conversations has a certificate parser (yes and so does almost everything that uses TLS)
It's trivial to use TLS without writing your own certificate parser. Doing this means taking on a lot of unnecessary risk, such as CVE-2023-33202.
Your encrypted messaging application shouldn't need to have a separate X.509 or ASN.1 parser built into it. If you're going to use them from TLS, you should rely on the library your OS vendor maintains for you, since they have an incentive to keep theirs secure anyway.
"Ominously pointing out" that the Conversations project has taken on an unhealthy amount of complexity and risk isn't FUD, it's a criticism of how the project is managed. Confuse the two at your own peril.
I use Conversations as a client in Android, and Gajim in Linux.
The basics like OMEMO, file attachments, push-to-talk voice memos (in Conversations, but not Gajim) work well. All comms are at least client-to-server-encrypted, and the OMEMO-protected comms are end-to-end encrypted.
Prosody allows voice and video calls - I have a TURN server all set up with it - but QOS rules (which are beyond my control, and are the fault of the ISPs I use) can sometimes squash calls made.
Gajim 1.9.3 (Flatpak) seems support voice and video calls to Conversations, but then it doesn't work. One Conversations smartphone can make voice and video calls to another Conversations smartphone.
I tried Matrix but found it to be more complicated without adding significant functionality.
This simply is untrue. Having first class E2EE that covers VOIP and all features (eg Matrix, Signal etc) like that is obviously a benefit.
Essentially you can't "use Matrix/Signal wrong" and realize later your communications weren't E2EE, you certainly can with XMPP.
It has gotten much better in recent times, but not quite there just yet.
So the real question is which one do you want to use?
In terms of server software, XMPP is lighter, easier to manage and scales better.
In terms of protocol, the Matrix protocol has an immense metadata problem. While encryption is sound, the metadata carried with every message is as good as plaintext. XMPP meanwhile provides pseudonymous rooms and minimal metadata leakage.
In terms of direction, I trust the XMPP foundation more. The Vector foundation comes from a multinational and is clearly a startup. I don't think anything good can come from the tight coupling between Vector and Matrix, no matter how much Matthew tries.
Vector also relicensed Element under a CLU (Contributor License Agreement). For me, it's a sword of Damocles.
In terms of community, it's pretty personal but I don't like the people on Matrix at all. The XMPP users seem more friendly. Your mileage may vary though. It may also not matter if you only want to talk to your already existing friends.
However, the best XMPP client is beautiful like the face of a dying man, and just as easy to use. Matrix is way ahead on client design.
You will need to pick a client that you can recommend to your friends and family, because they likely won't bother to look one up. If they are non-technical, then the client is what matters most for them and you need to decide what makes the most sense given their circumstances.
On the matrix side I can recommend Element (https://element.io/) which has a lot of eye candy. I'm not sure what to recommend for XMPP.
You also need to pick a server for them, because again, they won't bother to choose one themselves. If you can, host a XMPP or matrix server yourself!
As long as you stayed on the same instance everything was dandy, but writing to someone one the matrix homeserver worked most of the time.
Matrix as a protocol is more exciting than xmpp, but using xmpp (without omemo unless you are all on Conversations) is boring in a good way.
We've seen the appointment of people with biographies suggesting affinity with US interests, for example Katherine Maher to the Signal Foundation (and the departure of the Moxie Marlinspike.)
I see a members list for the board overseeing the spec for Matrix, but does anyone know who they are[1]? And the spec is one thing, but who controls the development of the actually used clients and implementations?
Please note I am not imputing anything w.r.t. Matrix/Element etc, but if we're bothered to put effort into E2EE then it is probably worth thinking about this aspect in addition to whether the technical basis is sound. The historical record is clear on government attempts to influence things from that end.
I have the same sort of worries about GnuPG (and the SPoF that is the hard-working maintainer.)
Maybe E2EE is not of concern to the OP, in which case disregard the above.
My issue with anything not Synapse (for now), is migration and atability. Conduit is still somewhat a young project in comparison, and I'm not entirely confident in the upgrade paths yet. They also use RocksDB (or sled), which furthers this "fear", perhaps irrationally so. I'd love to host my conduit server with confidence though, but I'd also need Matrix to properly support account migration. IMO it's one of the key features still missing and it's damaging its "decentralized" image.
The desktop client is capable but very very janky.