sanjjaystars

ThinkStick launcher. Installs a command that fetches a portable llama.cpp engine plus a quantized GGUF model on first run, then serves a fully offline local AI assistant on 127.0.0.1. Nothing is ever sent off the machine.

Install

pip install sanjjaystars

Installing does not download or run anything — it only puts the sanjjaystars command on your PATH. That's intentional: packages that silently download-and-execute at pip install time are exactly the pattern PyPI's security scanners flag as a supply-chain risk, so this project keeps install and run as two separate, explicit steps.

Run

sanjjaystars run

First run downloads the engine (~100 MB) and the default model (~4.7 GB, Qwen2.5-Coder) into a local cache directory, then starts the server and opens your browser. Every run after that reuses the cache and starts in a few seconds.

sanjjaystars list-models          # see what's available
sanjjaystars run --model research # switch models
sanjjaystars run --model math
sanjjaystars download             # fetch without starting the server
sanjjaystars run --force-download # re-fetch even if cached

Cache location defaults to ~/.cache/sanjjaystars (Linux/macOS) or %LOCALAPPDATA%\sanjjaystars (Windows). Override with the SANJJAYSTARS_HOME environment variable.

Before you publish this to PyPI

The download URLs in src/sanjjaystars/config.py are pinned to specific llama.cpp release tags and specific GGUF filenames. Both go stale — llama.cpp cuts new releases regularly, and quantized model repos on Hugging Face get renamed or re-uploaded. Verify every URL in config.py against the live release/repo before shipping, and budget for updating them again on release day. If a user's download 404s, that's the first place to check.

License

MIT — see LICENSE.