Tulir Asokan / Blog

Project updates (and a blog)

Posted on • 1653 words

I haven't been posting on TWIM for a long time and too many things have piled up to post everything in TWIM. This post recaps the potentially interesting things that I've worked on in the past ~half a year.

Also, since Automattic is basically a blogging company, I decided to try making a blog. Can't promise there'll be any more posts though, last time I got to a grand total of three.

Next generation of mautrix bridges

mautrix bridges have historically involved a lot of duplicate code across the different bridges. Each bridge has had a very similar structure, but there were lots of small network-specific differences, which made it difficult to share code. This also made it difficult to build new bridges without having extensive knowledge of how the existing bridges worked.

Recently, Beeper decided to move all bridges to run in clients instead of the cloud. Client-side bridges have lots of similarities to cloud bridges, but they also have some differences, quite similar to how different networks have slight differences. Due to the differences, our initial proof-of-concept local signal bridge was only able to reuse the message conversion code and none of the other bridge code.

Introducing Megabridge

To solve both problems at once, we started the megabridge project. The primary output is the new bridgev2 module in mautrix-go, which contains all the generic bridging code and defines interfaces for both remote network connectors and Matrix connectors.

The name "megabridge" reflects the additional possibilities enabled by the interface approach: one could build a binary that runs multiple bridges in one process simply by creating multiple instances of the relevant interfaces. In the old architecture, bridges were strictly singletons. Beeper also has a project called "megahungry", which applies the same idea to hungryserv, our unfederated Matrix homeserver. Beeper runs hungryserv and bridges in a single-tenant model, but the new megahungry and megabridge projects will allow running multiple single-tenant instances in one process.

Matrix connectors

Two Matrix connectors have been written: one inside the Beeper app for local bridges, and another in mautrix-go for connecting to a standard Matrix server as an appservice. When using the appservice connector, the bridge will work exactly like existing bridges do.

In the future, a third connector could use MSC4144 instead of an appservice, enabling self-hosting bridges against any homeserver. MSC4144 allows a bot to specify the displayname and avatar separately for each message, which is the main blocker for running bridges using a single bot account. There would still be some limitations compared to appservices, e.g. you wouldn't be able to view the remote member list or get read receipts, but it's better than nothing.

Network connectors

The current bridges will be replaced by implementations of the network connector interface, which effectively means all existing bridges will be rewritten. The Signal bridge already includes a v2 connector with all the basic functionality.

Rewrites of the Python bridges (Telegram, Twitter, Google Chat, LinkedIn) are underway, and the other Go bridges will follow later. I'd expect somewhat functional network connectors for all networks within a few months, although achieving feature parity, especially for features that aren't used in Beeper, will take longer.

Writing new bridges

The interfaces are still evolving, but the current features are already functional. If you want to build new bridges, it should already be much easier to get started with the new architecture than the old one. See the unorganized-docs directory in mautrix-go for more details and docs. The docs will eventually be organized into docs.mau.fi and/or pkg.go.dev. You can also join the #go:maunium.net room to ask questions.

At some point in the near-ish future I'll try writing a bridge to a simple network (perhaps Twilio) as an example and make a blog post showcasing the process.

Old bridges


The Go rewrite of mautrix-signal was finally released in February after being in development for several months. The old bridge is fairly broken by now due to Signal API changes (e.g. signald can't link new devices at all anymore), so I assume most people have already updated.


The legacy Facebook and Instagram bridges were replaced by a new Meta bridge which uses the web app API. The API happens to be pretty much the same for both FB and IG, which is why the bridges were merged. The old bridges may still work to some extent, but they're not being maintained anymore, so they'll break more and more as Meta stops supporting the old APIs.


After Beeper joined Automattic, we had to start using Slack when talking to Automattic people (we still use Matrix/Beeper for internal chats), which means I had to actually care about the Slack bridge.

I was planning on refactoring the bridge to bring it up to the level of the other Go bridges, but unfortunately didn't manage to finish the refactor before starting the megabridge project. At this point, it is unlikely that the refactor would be finished before the bridge is rewritten to use the new architecture.

mautrix-discord is similarly stuck on a slightly older mautrix-go version, so it probably won't get any more releases until the megabridge rewrite.


With all bridges moving to Go, mautrix-python has seen less activity. It will still be maintained, but the bridge module will likely be deprecated or even removed once the Go rewrites are ready. maubot still uses mautrix-python and is not going to be rewritten in Go. In fact, the original version of maubot was written in Go, but it was rewritten in Python due to the lack of easily reloadable plugins. As far as I know, the plugin package of Go has not improved much since then.

High-level client framework

I've recently started developing hicli, an opinionated high-level module for building clients. It manages room timeline storage, handles everything related to end-to-end encryption, and more. The storage layer already works to some extent, but more complex features like unread counts are not yet implemented. Additionally, it hasn't been tested in an actual client yet. Since it's a side project, it'll probably take a while to produce anything actually useful.

There are two reasons for this project. The first is that due to the atrocious performance of Element Web, I've wanted to make my own web client for a while now. Making a new Matrix SDK in JavaScript is a non-starter, so instead I looked into ways to use Go. Unfortunately, running SQLite in WASM with OPFS doesn't seem performant enough currently, so a pure web client is not feasible yet. Instead, I plan to write a separate daemon that a web client can communicate with over local HTTP, or perhaps through a webextension. The web client and daemon could also be bundled into a desktop app using Wails.

The other reason is that gomuks has been neglected for quite a while, primarily because I don't actually use terminal clients, but also because the data layer was grown organically and is rather bad. If I can make a good data layer for my web client, I can hopefully also rewrite gomuks on top of that and maintain both more actively.

Based on my experience of building an unfederated Matrix homeserver, SQLite is very powerful and extremely fast if used correctly. If it can handle being the database for a homeserver, it can certainly handle a client. To avoid some of the mistakes in gomuks, hicli uses SQLite for storing everything and even uses some of the more advanced SQL features like triggers.

Gifs in maunium-stickerpicker

The sticker picker is more or less ready. It could still use a server component for easy multi-user setup, but I'd hope that native stickers will happen within a year or two, so spending a lot of extra effort on the sticker picker isn't worth it. However, @hedgenischay recently contributed support for sending gifs via Giphy, which means Element Web can now have a real gif picker.

An interesting detail of the gif picker is that it doesn't reupload gifs. Instead, it simply sends mxc://giphy.mau.dev/* URIs, which redirect to i.giphy.com. More on that in the direct media section below.

The feature is enabled by default in the latest version and doesn't require any extra configuration, although it is possible to override the gif proxy and giphy API key if desired.

Direct media access

mautrix-discord has had direct media access for a while now. The general idea is that instead of reuploading media, the bridge will generate a mxc:// URI which actually redirects to the Discord CDN. This means less wasted disk space, and if your homeserver caches it as remote media, it's easy to clean up.

The original implementation of direct media was simple enough that the redirects could be implemented inside a reverse proxy. However, around February, Discord started using expiring signed download links, which prevented a pure reverse proxy implementation. A new version of direct media was built into the bridge to support refreshing the download link if it expires. The new version also had extra fun stuff, like implementing the /_matrix/key/* endpoints to pass the federation tester.

More recently, MSC3916 was accepted to add authentication to all media download endpoints. The original implementation would've blocked direct media redirects, but a late change shortly before accepting re-added proper redirect support.

Even though it allows redirects, MSC3916 still makes the federation download endpoint more complicated, because the response must be a multipart/mixed. This makes it harder to implement purely in a reverse proxy, so a small proxy service is usually required. To make it easier to make such small services, I moved the direct media code from mautrix-discord to mautrix-go: https://github.com/mautrix/go/blob/master/mediaproxy/mediaproxy.go.

The giphy.mau.dev proxy mentioned in the sticker picker section uses the new media proxy module in mautrix-go: https://github.com/maunium/stickerpicker/blob/master/giphyproxy/main.go.

In addition to mautrix-discord and the sticker/gif picker, the bridgev2 project will include an interface to let network connectors implement direct media access using the mediaproxy module.