I kept spending my money on food instead of upgrades. Here’s hoping DDR5 drops below 2x its Sept 2025 value in 2027, to coincide with Zen 6.

I tried to get RAM on eBay. 8GB & 16GB kits are not even being auctioned, and when they are sold their price tag is well over 3x their value, often 4x (~GB£170). 32GB sells for 3-4x its value (~£260-£320), almost as much as 64GB, which is often going for >2.5x (~£380-£440). If you’re desperate or are okay with massive financial waste, eBay is often the cheapest place for 64GB kits, but anything lower is relatively more expensive than just buying from distributors

It was a huge pain and I ended up troubleshooting with Gemini for hours aha! I know, I’ll plant a tree to offset my sins. It was at least useful to rapid search solutions and tell me what component was the most likely issue.

I had coturn set up for legacy Element Classic and, before that, XMPP, but as I wasn’t using those I decided to shut it down and try using Matrix Livekit’s internal TURN server. I’m not sure what actually helped in the end, but Livekit’s latest build caused a bug, so I instead pulled v1.9.12. I also shuffled around my reverse proxy config (from my old attempts) because some endpoints seemed to have changed. I’ll update later with anonymised config :3

Hey, just coming back to see how your setup’s going, and to say I’ve finally managed to get Element Call working for Matrix – I can help you get it running if you like!

I like that. The machine I use to host ytdl-sub is called ourtube

Mad props to the dev for a GUI

It is indeed rather complex.

I think you got a downvote for promoting mailcow, users can be fickle

This is good, I use Mailcow Dockerized and it uses 10% of one 3.7GHz core, but 2GB of RAM. Stalwart definitely seems better for low memory hosts. Seems to have one instance of rspamd for each mailbox, they and ofelia are the biggest users of RAM according to top

Can confirm what another user said, that Intel iGPU would be better in your case.

I’ll let you know now – if it runs Windows kill it. My server was originally Windows running Docker Desktop. It hosted three services: Minecraft server which lagged like a bitch; Samba folder share; and Emby. Whenever Emby playback froze I knew Windows, whose antivirus kept running the HDD under constant load, had fucked the i6 6100 to 100%, which happened at least twice a day.

Moving on, now I run Proxmox. I host 25 services with the CPU at ~35% idle and 24GB RAM at 75%. Nothing lags.

Before I plugged in the GPU my server drew 25W consistently, going to 35W under load. With the GPU, an RTX 3060 11GB (used), it uses 85W idle, so make sure it’s worth it. For my case it not only transcodes for Emby and resumes streaming in a second, but also handles voice inference for Home Assistant in under a second, and mid-sized Ollama LLM responses. Would recommend a high VRAM Nvidia card (for CUDA) in that scenario, as my model Gemma3 7B uses 6GB VRAM and 2GB RAM. But a top model, say Dolphin-Mixtral 22B, needs 80GB storage, 17GB RAM and… Well I don’t have the RAM but you get it. LLMs are intensive.

Here you are :) it’s a Github link (I’m looking into hosting a private pastebin like PrivateBin and will replace this later)

My install must have been broken then 😭 and my experience is from around early 2025, and I didn’t keep it around, so my intel is also dated…

Ooh they upgraded? Yeah my information is based on early 2025 when I tried it aha

Oh, and if you wish, it’s a bit old now but no doubt useful, I have written installation guides on both Prosody and Continuwuity, based on Proxmox containers.

As someone who’s tried both, it depends on what you want. Your choice of Matrix server depend on any political and ethical values – some say Synapse is too corporate, being maintained by Element who are for-profit and obtain funding from corps and governments, so some prefer others such as Conduit ( – until maintaining slowed to near abandonment and it was superseded by Conduwuit – until the owner got cyberbullied so hard she quit the project and it was superseded by Continuwuity) because it was built on Rust and much more efficient than Synapse, or Dendrite. I recommend Continuwuity.

Then there’s clients – the only mature matrix client for mobiles is Element, and there are two apps, Classic and X, who offer different pros and cons, and imo are not good enough on their own, both are in a kind of beta stasis. But it’s the best they have. If you really don’t need calling, then Element X, FluffyChat or Schildichat is your app and Element Web for desktop access (available on Github). However, when exchanging encryption keys to trust another of your devices, or a contact’s device, only Element offers simple QR scanning.

In short, Matrix is very good as a privacy-focused server with partially working, modern looking clients.

Then there is XMPP. Again there are different backends to choose from and I am inclined to recommend Prosody. XMPP just works out of the box for me, calling included, and is relatively stable. However, there are large caveats – several pieces of user data are stored unencrypted on the server, which is fine for you as the owner, but it’s a lot harder for someone else using your service to trust that. And, while XMPP uses OMEMO encryption keys, handshaking with devices is far more manual than Matrix’s Olm/Megolm and involves a multi-step process, and migrating to a new device is a pain because messages are not backwards decrypted, so they must be transferred from the first device. Finally, clients are very rough. The best desktop clients such as Gajim and Pidgin still look like they were built in 2001, and while mobiles have Monocles, Cheogram and Conversations, they all look very similar, as the former are very slight modifications of Conversations.

In short, XMPP may lack some comforts of modern messengers, but it is simpler to set up than Matrix, and offers many of the same features. However, the manual key sharing process might scare off all but the most avid privacy enthusiasts, especially that if you migrate to a new device without sharing message history from a previous verified device, messages are lost.

Choose Matrix for polished software, inviting many contacts, and, with Element X featuring (eventually) Element Call, complete E2EE.

Choose ol’ faithful XMPP for an easier initial setup, if video calls are important, you appreciate that historical messages cannot, by design, be hacked into, or if you don’t like Element the company.

I too have heard good things about SimpleX and Signal, and recommend trying them if they are valid contenders for your use case. Signal really is the best (most private, least data-farming) non-selfhosted option.

I don’t think there’s a straightforward way like a HACS integration yet, but you can access Ollama from the web with open-webui and save the page to your homepage:

Just be warned, you’ll need a lot of resources depending on which model you choose and its parameter count (4B, 7B etc) – Gemma3 4B uses around 3GB storage, 0.5GB RAM and 4GB of VRAM to respond. It’s a compromise as I can’t get replacement RAM, and tends to be wildly inaccurate with large responses. The one I’d rather use, Dolphin-Mixtral 22B, takes 80GB storage and 17GB min RAM, the latter of which I can’t afford to take from my other services.