Built a self-hosted control panel for local AI on your own GPU box. Run llama.cpp chat, stable-diffusion.cpp images, and audio.cpp speech/music from one browser tab; Optimizes, benchmarks, fits, abliterates, scales, and educates. DavidAU qwen3.8-27b guffs as baseline, reproducable quants at speed.
retained text
\*\*Edit\*\* Thanks for the award! Built a self-hosted control panel for local AI on your own GPU box. Run llama.cpp chat, stable-diffusion.cpp images, and audio.cpp speech/music from one browser tab. Yes it is vibe-coded, It's named after my kid 'Panel', and I made this for me because everything else sucked or was too secretive. Constructive criticism appreciated. Still refining and love good ideas. [https://github.com/W61k3r/LexiPanel](https://github.com/W61k3r/LexiPanel) What it does: \- One panel per GPU — separate params, ports, logs, systemd units \- 200+ settings with plain-English tooltips (no guessing) \- Preview exactly what would launch + VRAM/RAM estimates before you start \- Fit engine calculates how big models fit across your GPUs before loading them \- Power options: CPU governors, GPU caps, fan curves, PCIe/NVMe power saving \- LoRA adapter support (abliteration/uncensoring) \- Built-in web UI integration — toggle on/off, set defaults (system message, theme, reasoning display), protect behind the panel login via Caddy \- Daily build updater for all engines \- Benchmarks, crash triage, file manager, web terminal \- Graph Gauntlet mini-game on every chart (I got tired of staring at graphs) Pure Python — stdlib only, zero pip deps, no Docker, no database. Built to optimize and stabilize 7900xtx build, works for nvidia/amd/etc though.
document fetched 2026-09-25 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1wp9kpu/built_a_selfhosted_control_panel_for_local_ai_on