Simpler self hosted alt to Open WebUI
Posted by anitamaxwynnn69@reddit | LocalLLaMA | View on Reddit | 17 comments
Got Qwen3.6 27B running on my newly assembled 4x 3090 rig (s/o 3090-club) and I'm trying to get the people in my house to adopt the local workflow.
Open WebUI has improved a lot in the recent updates, but I still found it pretty rough for non-technical people. It often feels more like a dev tool than a self-hosted ChatGPT-style app that "just works". I built overtchat to focus mainly on getting the core chat experience right: a polished ui, simple setup and fewer moving parts. The goal is not to compete on agentic workflow with LibreChat/LobeChat/OWUI but to provide a cleaner self-hosted interface for local models.
Ships with its own tried & tested searxng config for web search, kokoro tts (no api keys needed). Single docker compose file. MIT licensed of course, no telemetry. Optimized for mobile as PWA. Github.
Also being upfront - I write code for a living and have been actively reviewing/debugging/changing things, but I did use quite a lot of AI lol. I promise it's not slop tho 😿 . Feedback is welcome!
kevinlch@reddit
if you're using llama.cpp they have a pretty decent ui too for simple chat like this.
anitamaxwynnn69@reddit (OP)
Yeah, llama.cpp’s web UI is great for a lightweight UI close to the inference server. OvertChat is aimed more at the self-host this for the household use case: users/auth, first-user admin, LAN-friendly setup, and mobile/PWA usage. Basically less single interface to llama-server and more ChatGPT-ish app I can give to family/partner who don't need to know the stack lol
Opening-Broccoli9190@reddit
llama.cpp is extremely easy to use without knowing the stack, what seems to be the hiccup with your partner?
emiliobay@reddit
Local LLM UIs often treat STT as an afterthought, but it is actually the biggest adoption blocker for non-technical users. The model and TTS can be flawless, but if household members have to hold down an awkward key combo or leave a hot mic running, they will just stop using it. Moving the voice trigger off the keyboard entirely to a dedicated push-to-talk button completely changes how approachable the system feels.
pl2303@reddit
Looks good, but Kokoro is limited regarding supported languages, so I have to pass on it.
Otherwise_Economy576@reddit
oh nice. how does the multi-user thing work in practice? the bit where 'people in my house adopt the local workflow' is usually where it falls apart for me. owui has SSO/oauth but it's overkill for a household. did you go with simple password auth or something else?
anitamaxwynnn69@reddit (OP)
email+pw for now, better auth under the hood. first signup goes to the admin and then locks the sign up page; after that the admin adds accounts from settings (temp pw, user changes after first login). no sso/oauth/invite emails. spin it up once, add 3-4 users, that's it. the bigger adoption question for households isn't auth haha, it's whether the alt "just works" so they can feel comfortably using it instead of chatgpt :')
Ha_Deal_5079@reddit
4x 3090 gang lets go lmao. hows the power draw under load? and yeah i feel u on open webui being too much for just chatting
anitamaxwynnn69@reddit (OP)
1300-1400W at peak but that's the worst case including CPU/DRAM. I've already seen discussions on this subreddit about power limiting I'm planning to do an experiment to find the "sweet spot" for me. Found this article (not mine) and I plan to conduct a similar one once all my pcie risers arrive. Right now they sit too close to each other and the vram will literally go to 110 and I do not intend to operate it like that haha.
ego100trique@reddit
Try to undervolt them firstly then power limit, should be helpful for power draw and coil whine
wayofTzu@reddit
Very interesting! My biggest gripe with OWUI is the persistent storing of information (e.g., voice files, chats stay on disk forever without a custom script). Can I ask, how does this handle such things? TTS, does it save the file or convert it and dump it? File uploads, are they assessed then dumped or also stored?
anitamaxwynnn69@reddit (OP)
Yeah, that’s a fair concern and something I care about too. TTS is one-shot right now, it streams the audio and nothing is stored. Uploads are stored right now so chats can keep referencing them later, but abandoned uploads that never make it into a message are cleaned up after 24h. Chats are persistent until deleted at the moment. I don’t have auto-retention yet, but I’ll be adding that in the next update along with temporary chats.
Let me know if you have any other feature requests!
wayofTzu@reddit
Exciting work, thank you for making it so readily accessible! I'll be giving it a try in the future. I have a similar use case as you: Just want to share simple versions of my tools with house mates.
o0genesis0o@reddit
You know, with all the supply chain attacks happening lately, and the capabilities of coding agents, it feels like we will have a very different view about software and software engineering in near future. Sometimes I feel it's easier to have a nice scaffolding that I trust, and then give LLM the right skill, and build out whatever I want from there than hunting for obscure package on github.
Btw, have you tried ddgs library? I used it instead of hosting my own searxng to enable web search. Seems to be working alright with no infrastructure overhead.
anitamaxwynnn69@reddit (OP)
Yeah I think that’s fair. I’m trying to keep the stack pretty boring (battle tested) for exactly that reason fewer moving parts, no plugin/RAG/vector DB layer, etc.
I haven’t tried ddgs yet but I’ll check it out. I went with bundled SearXNG mainly because I wanted the search path to stay fully self-hosted and predictable, but if ddgs works well enough with less infra, i can look into it!
slavik-dev@reddit
Nice.Â
I see TTS included. What about STT? Is it planned?
anitamaxwynnn69@reddit (OP)
Yes, I'm planning to add ParakeetV3 0.6B next. It is faster and more accurate than whisper. I'm open to suggestions though!