Pi Coding Agent + Docker Sandbox + llama-server: Setup Guide
A setup guide for running the Pi coding agent inside an isolated Docker Sandbox microVM while it talks to a local llama-server inference backend running on the host GPU.
This repository is a guide for combining three pieces: the Pi coding agent, Docker's sbx sandboxing tool, and llama.cpp's llama-server. Pi runs inside a hypervisor-isolated microVM so it cannot read host files outside its mounted workspace, cannot reach the host's keychain or SSH keys, and cannot make network requests beyond an explicit allow-list.
Inference is served by llama-server running in router mode on the host, so models load dynamically and run at full host-GPU speed while Pi itself stays sandboxed; requests cross the VM boundary through the sbx proxy's host-gateway address. The guide walks through installing sbx, installing or building llama.cpp with CUDA or ROCm acceleration, and the remaining setup steps.