Image: reddit.com · Twissbot (+my AI experience so far)
TwissBot: a multi-agent local LLM assistant on old GPUs
A self-hosted chatbot vibe coded with Claude that splits questions across small local models to get quick answers from an old GTX 1060.
Credits
Earliest dated source:
The full story
TwissBot runs on a free HP ML360 Gen 8 server that came with a GTX 1060 6GB. The maker, who says they are good at setting up systems but terrible at code, vibe coded it with Claude and is on the fifth version.
A single agent was slow on the old card, so it now passes small prompts to different models: one breaks a question into smaller ones, and Qwen 2.5 7B, Qwen Coder 2.5 7B and Mistral 7B handle the intro, code, explanation and outro. The maker reports "really snappy" responses and plans a control agent handing work to a GTX 1080 worker card.
How it was done
Tools
Themes
Loading themes…
Sources
From a short post by the maker, so the details are as they tell them.
- Reddit: Twissbot (+my AI experience so far)Reddit poster: twiss2020Posted
The maker's post on r/homelab.
If it disappears: archived copy of Reddit: Twissbot (+my AI experience so far) (opens in a new tab)