muthr 0.1.23

muthr is a zero-trust orchestrator that automates llama.cpp and Lima to run local AI agents. It controls inference via a host-based llama-server and spawns isolated Lima VMs for agent execution. Agents get full read-write access to your project workspace, but zero access to the host OS or SSH keys.

Architecture

1. llama-server on macOS, accelerated via Metal
2. limactl VMs provisioned per-project
3. opencode inside guest VMs, connecting over host.lima.internal

Prerequisites

macOS (Apple Silicon, ≥48GB RAM for 35B models), Lima, llama.cpp

[!NOTE] The ≥48GB RAM requirement applies to 35B models. Smaller models run on machines with less memory.

Usage

muthr init               # Clone configs from tappunk/muthr-configs
muthr download <source>  # Fetch GGUF model from HuggingFace

muthr serve              # Start llama-server as a background daemon
muthr serve --foreground # Run in foreground
muthr status             # Check engine status and active profile
muthr stop               # Stop the engine

muthr up                 # Provision a Debian 13 VM for the current project
muthr ls                 # List all active sandboxes
muthr down               # Stop the current sandbox

muthr services start     # Launch MCP services VM
muthr services status
muthr services stop

Configuration

Config in ~/.config/muthr/ (see muthr-configs for the full directory structure and examples):

• provider.d/llama-cpp/*.ini — preset profiles (context sizes, threading, model paths)
• clients/opencode-config.json — template for OpenCode runtime config generation
• manifests/*.yaml — VM architecture, memory, container configs
• provision.d/*.sh — boot scripts for OpenCode CLI and dependencies

Runtime state (PID files, logs, generated JSON) in ~/.cache/muthr/.

Installation

muthr is available on crates.io and Homebrew.

Cargo

cargo install muthr

Homebrew

brew install tappunk/muthr/muthr