hostely open-sourced: C++20 Apple-native CLI with honest unified-memory reporting

ayo_ham · x · 2026-09-06

The hostely repo is now public: written in C++20, MIT-licensed, built with CMake, with a quickstart and smoke tests in the README. The author especially wants feedback from M1/M2 users with less RAM.

The GitHub page showcases the core selling point — honest unified-memory reporting: hostely status shows total memory (32GB on the demo M4), current usage, and computes "model headroom" (e.g. 11.7GB = free + inactive − 2GiB safety margin) plus a Metal recommended max (21.3GB), refusing to load models that won't fit. The tool bundles Apple's container runtime for OCI containers and a Metal-accelerated llama.cpp server for GGUF models, plus HTTPS, DNS and tunnels — no Docker, no ngrok. A 1.1B TinyLlama Q4KM peaks at just 980MB in the example.

Related event: hostely: Open-Source CLI Brings Containers and Local LLMs to Mac mini(3 posts)→

Original post →

More from coding & agent

coding & agent channel →