dotfiles

 # Secure AI-Workstation > **A local AI stack that mitigates supply chain attacks and misbehaving agents.** **⚠️ Important Warning** This is a **security research project** and my personal dotfiles. It significantly reduces the supply chain attack surface while delivering excellent local AI performance on Apple Silicon through overlooked security best practices. **Only proceed if you are comfortable with:** - Declarative Nix systems and `nix-darwin` - Managing system-level configurations - Reviewing code before execution --- ## Features - **Agent tool-call isolation** — Each project runs in its own reusable Lima VM with **sandboxed OpenCode agents** that have full read/write access to the project directory but are isolated from host SSH keys, home directory, and other projects. - **Third-party service containment** — Dedicated Lima VM for MCP services (e.g. SearXNG) to isolate external dependencies. - **Auditable inference engine** — `llama.cpp` built from pinned source with Metal acceleration. Dual-model support (primary + utility) via presets. - **Reproducible dependency chain** — Everything managed via Nix with SHA256-pinned hashes. No runtime downloads. - **Hardened dev environment** — Ghostty + Neovim (full LSP) + Starship + Lazygit. --- ## Architecture ### Execution Environment Debian 13 VMs via Lima (`vmType: vz`) running **sandboxed OpenCode agents**. ### Inference Engine Single `llama-server` instance serving multiple models simultaneously. ### Tooling & Virtualization Nix-managed tools with Apple’s Virtualization.framework for high performance and low overhead. --- ## Hardware Requirements - **Recommended**: Apple Silicon with ≥48 GB RAM for smooth dual-model inference. High memory bandwidth is critical. --- ## Quick Start ```bash # 0. Install Ghostty (recommended) # https://ghostty.org/download # 1. Xcode Command Line Tools xcode-select --install # 2. Determinate Nix curl --proto '=https' --tlsv1.2 -sSf -L https://install.determinate.systems/nix | sh -s -- install . /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh # 3. Bootstrap git clone https://github.com/tappunk/dotfiles.git ~/dotfiles cd ~/dotfiles ./bootstrap.sh exec zsh # Rebuild anytime system-rebase ``` > First run compiles `llama.cpp` from source (safe to ignore `$HOME` ownership warning). ### Download Models ```bash llama-download https://huggingface.co/.../Qwen3.6-35B-A3B-UD-Q4_K_M.gguf llama-download https://huggingface.co/.../Qwen3.5-9B-Q4_K_M.gguf ``` Then run `llama-serve`. ### Launch Sandbox ```bash sandbox-start # run from inside your project folder ``` ### MCP Services VM (optional) ```bash mcp-services-vm-start ``` - SearXNG: http://127.0.0.1:18766 - MCP API: http://127.0.0.1:18765/mcp