Another AI slop..
I won't argue with the downvotes, honestly they're fair and I get it. This is an AI project posted to a community that's rightly tired of AI slop, and I'd be skeptical too.
So let me be straight: yes, I used Claude heavily to build this. I'm a solo dev and it's how I got it done at all. But every decision (architecture, stack, features, scope) is mine. That doesn't mean they're good decisions, just that they're deliberate, not generated. I built this because it actually solves a need I had at home, and I figured I'd share it in case it's useful to someone else, not to pretend I'm reinventing anything.
I also added an honest "How this is built" note to the README and the site rather than hide the AI use. If the code reads like unreviewed slop anywhere, that's a real bug to me and I'll fix it.
Either way, I hear the reception loud and clear, and I appreciate the people who took the time to actually tell me why.
This looks interesting, especially the persistent memory. I want to try it out but it seems likely to me that multiple simultaneous agents would require significant hardware. Even if they were serially activated, reloading contexts with each switch would take time. I have a pretty beefy GPU and experience significant (almost ridiculous) slowdown when opencode runs 2 subagents simultaneously.
But perhaps the memory storage/lookup keeps contexts very small?
Anyway, I can't find any mention in the repo or docs what the suggested minimum hardware is.
Good question, and you're right that it's missing from the docs (just added a Hardware requirements section to the README to fix that).
The key thing: Hivekeep doesn't run the models itself. It calls your provider (Anthropic/OpenAI/etc.) or a local OpenAI-compatible endpoint, so the heavy compute lives there, not in the app. The platform itself is a single Bun process over SQLite, no GPU, no extra services. It runs in well under 1 GB of RAM on a small home server.
On the multi-agent worry: agents are activated serially per message, not all firing at once, and the persistent memory is exactly what keeps each context small (hybrid vector + keyword recall, re-ranked, instead of replaying the whole history). So adding agents mostly means more calls routed to your provider, not multiplied local load.
The opencode slowdown you saw is on the inference side: if you point Hivekeep at local models (llama.cpp / LM Studio / Ollama / vLLM), the hardware question moves to your inference server, same as any other client. If you use a hosted provider, your machine barely feels it.
Yeah, my thinking was definitely biased to my environment. I selfhost llama.cpp so even if Hivekeep doesn't require significant resources, whatever LLMs it runs will use my hardware.
agents are activated serially per message, not all firing at once, and the persistent memory is exactly what keeps each context small
It sounds like I need to try it!
Thanks, that means a lot! And I'm not sure if it was you who sent me a feedback through Hivekeep's in-app feedback system, but if it was: thank you, it was genuinely constructive and I'm going to work on reliability with lighter / self-hosted models.
To your question: yes, that's basically the core idea. Agents build their own tools and mini-apps (small interactive UIs embedded right in the chat) from inside the platform, so you can grow it toward the assistant features you need instead of waiting for them to ship. I'm not familiar with Odysseus specifically, what did you have in mind? Happy to tell you if it's doable today or where it would land.
I haven't found the time to try it yet, so it wasn't me the one who used the in-app feedback feedback system. But what I have in mind is: pull emails, Forgejo issues, calendar, etc, show me a kanban and Eisenhower matrix of what I have to do or to delegate to either agents or other people.
thanks for using Leebra!
go to feed...
Friendly feedback (genuinely) per your request.
Cardinal sins on Lemmy -
Speaking of, code base itself has "llm wrote this, no one checked" stank.
I pointed Claude at the first file I could see
https://github.com/...
Issues identified (first 5, trimmed for brevity).
Ghost import (
TEST_USER) - classic back-fill residue. Pulled in because it "felt right" alongside the other two helpers, never wired up. No fail-loud catch on it - lint should've flagged it, didn't.Copy-paste boilerplate x3 instead of
beforeEach- pattern got nailed once, then stamped two more times instead of abstracted.networkidlereflex - slapped after everygoto, despite Playwright's own docs actively discouraging that specifically.Hedge-bet regex (
/password/i+ filter on/match|mismatch/i). LLM didn't know the exact copy, didn't go check it against source (the grounding step), so it just widened the net to pass either way. Not minding that it hallucinates, in test form.Comments narrating the line directly below them - zero information gain, pure token filler.
= you're going to get downvoted to hell because this stinks of AI slop.
Personally, while Lemmy is militantly FuckAI, I think properly orchestrated, ticketed and reviewed AI code is defensible.
https://lemmy.ml/post/48724623/26190950
AI slop (which is sadly what this reads like - please prove us wrong) is going to get kick fucked to death...especially on !selfhosted.
Not trying to dunk on you...trying to give you a chance to tell us why we should care about this project / install it / what problem it solves.
save