{inercia}

Link: chad and HakoVM — shrinking the agent stack down to one Mac

Link: chad and HakoVM — shrinking the agent stack down to one Mac ​

A couple of interesting AI-agent infrastructure projects worth investigating. Both projects strip a general-purpose agent-infra layer. One shrinks the model (chad), the other shrinks the isolation boundary (HakoVM).

  • chad (https://github.com/nathansutton/chad): a local Claude-Code-style coding agent running entirely on Apple Silicon via MLX, one 27B model (Qwen3.8-27B, UD-Q3_K_XL quant), no API key. Fitted engine (DFlash2 speculative decoding + persistent prefix KV cache + fused Metal kernels) gets 62 tok/s decode vs 10.9 tok/s on stock llama.cpp/Ollama, same M4 Pro (24 GB). Deliberately narrow: five tools, no model picker.

  • HakoVM (https://github.com/las7/HakoVM): Firecracker-class per-sandbox microVMs for macOS agent sandboxes, built on Apple's Containerization framework. Boots a hardware-isolated Linux VM from any OCI image in ~300ms, execs at 8ms p50 / 12ms p90 (n=20), no daemon/root, no NIC by default. Purpose-built for agent workloads: Claude Code Bash hook, MCP server, commit/clone deps workflow.

© 2026 {inercia}