🖥️ Own Your AI — a Private, Offline ChatGPT You Install, Not Rent

Tools & Scripts
, ,
1

:cloud: ─ ✕ ─ ─ ─ :play_button: :desktop_computer:

OFF THE CLOUD — your AI runs on your own metal · 0 cloud · 0 keys · 0 subs

HELIX chat dashboard
HELIX chat dashboard1482×992 120 KB

HELIX turns your own PC into a private ChatGPT — install it, pull any open model straight off HuggingFace from inside the app, and chat 100% offline. No account, no API key, no subscription, nothing ever leaves your machine. Built in Rust + Tauri, source on GitHub.

The LM Studio everyone runs is a closed binary. HELIX is the one you can open — read the source, run it on any GPU (Vulkan, not just NVIDIA), with agent plugins built in.

:down_arrow: Grab it — Windows

file size
HELIX-Setup.exe 11 MB one-click installer
HELIX .msi 18 MB enterprise/silent install
verify — SHA256SUMS.txt on the release

:link: Source + releases → github.com/tjcrims0nx/Helix

:green_circle: First chat in 3 clicks (zero skill)

  1. Install → open HELIX
  2. Search a model (try Qwen2.5-7B or Llama-3.2-3B) → one-click download (8-stream, auto-resumes)
  3. Type. Done — running on your GPU, fully offline.

:framed_picture: The tour

gguf
gguf992×836 72.1 KB

↳ Model Manager — search HuggingFace, parallel chunked pulls, live VRAM + disk analytics


mcp
mcp882×862 101 KB

↳ MCP Plugin Manager — hand your local model real tools: files, web APIs, agent skills

🔵 Under the hood — for the tinkerers
  • :gear: Vulkan + FlashAttention llama-server (-fa auto, ~3× prompt speed) — runs on Intel Arc / AMD, not CUDA-locked
  • :electric_plug: MCP plugins (stdio + SSE) via in-chat /mcp list · enable · disable · open → wire up filesystem / web / agent tools with an auto tool-call + reasoning-trace loop
  • :inbox_tray: Custom .gguf import · live model switching · real-time CPU/RAM/VRAM monitor
  • :hammer_and_wrench: Build from source — Tauri 2 (Rust) · Vue 3 + Vite · engine = llama.cpp
🍖 Feed it models — where the good GGUFs live
  • HuggingFace · GGUF trending — the firehose of open models
  • bartowski — cleanest quants of nearly every new model, days after release
  • :straight_ruler: quant rule of thumb: Q4_K_M is the sweet spot; rough VRAM ≈ (params-B × 0.7) GB → an 8B model fits in ~6 GB

:balance_scale: HELIX vs the usual

source GPU support agent tools
LM Studio :locked: closed binary CUDA / Metal MCP (recent)
Ollama terminal-first CUDA / Metal via 3rd-party UIs
HELIX :unlocked: open on GitHub Vulkan — any GPU built-in /mcp

Cloud AI rents you a seat and reads the room. This one you own — on your metal, behind your door.

1 Like