The Missing Cable Between Agents and Everything Else If 2025 was the year agents got good at writing code, 2026 is the year they got good at doing things — and the thing that made that jump possible is the Model Context Protocol. MCP is, at its heart, a boring and beautiful idea: a standard way for an AI agent to discover and call tools, regardless of who built the tool or the agent. ...
The Blog Name Was a Promise This blog is called The Least Privilege Life. I spend my days at Varonis helping organizations make sure attackers can’t touch data they have no business touching. And yet, if you scroll through my archive, you’ll find article after article about giving software more access — 650+ skills wired into my dotfiles, multi-agent orchestration, a modular AI forge. There’s a tension there, and it’s time I addressed it head on. ...
The Terminal is the New IDE In early 2026, the way we interact with code changed forever. I no longer spend most of my time typing characters; I spend it orchestrating context. My dotfiles have evolved from simple shell aliases to a comprehensive AI Orchestration Layer. This is the “Least Privilege Life” applied to engineering: providing exactly the right context to the right tool at the right time. The Ecosystem: 190+ Agents, 650+ Skills My environment isn’t just a shell; it’s a factory. I’ve integrated three primary AI CLI tools, all sharing a common brain managed through my dotfiles: ...
Beyond the Chatbot If you’ve followed this blog, you know my obsession with local-first systems. Running a basic LLM is easy. Building a professional-grade AI infrastructure that is reproducible, secure, and high-performance is a different story. Enter Synapse: my dedicated AI forge. While my Macs handle the interfaces, Synapse is the heavy lifter. Built on Fedora 43, it’s designed to provide a “zero-trust” environment for training and inference. The Hardware Stack To handle modern models (Llama 3, Mistral) with large contexts, I chose a balance of core count and VRAM: ...
If you self-host downloads with PyLoad, you’ve probably used Yape — the classic browser extension that lets you manage downloads without opening a web interface. It was simple, it worked, and then Google killed it. Manifest V2 was deprecated, and Yape was abandoned. So I rebuilt it from scratch. Yapee is the spiritual successor: a modern browser extension for Chrome, Firefox, and Edge that handles everything the old extension did, plus features that make managing multiple PyLoad servers actually pleasant. ...
The 32GB Wall My previous Homelab center—a custom-built i7 Proxmox node—was a champion for years. But as my development workflow shifted towards local AI, heavy container orchestration, and multi-OS security labs, 32GB of RAM became a suffocating bottleneck. I was constantly fighting swap death and micro-managing service uptimes. I needed a significant upgrade, but I wanted to stay true to the “Least Privilege Life” philosophy: maximum control, minimal noise, and efficient hardware reuse. ...
Let’s be honest: we all spend 90% of our lives in the terminal, and the other 10% wondering which of the 57 open tabs contains that one tail -f we started three hours ago. As a die-hard fan of Kitty (the fast, GPU-based terminal beast) and a daily user of Nushell (because why use old shells when you can have structured data?), I needed a way to bridge the gap between my macOS workflow and my terminal addiction. ...
I have a confession: I built an entire infrastructure to automate job applications. The technical achievement is worth more than any single outcome it produced. Here’s why I built it, how it works, and what it taught me about recruitment that you probably don’t want to know. The Problem You know that feeling when you’re applying to jobs and you’re manually tweaking your CV for every single position? Removing frameworks that aren’t relevant, reorganizing bullet points to highlight the right skills, adding specific keywords from the job description, rewriting your cover letter from scratch each time? ...
The moment you realize that a single LLM isn’t enough, you’re building multi-agent systems. You don’t choose to — you just end up there. I built HivePilot in late 2024, when multi-agent frameworks were mostly theoretical. The problem was simple: some tasks are too complex for one model to nail, but perfect for a swarm of specialized agents. Code review automation, documentation generation, integration testing — these need different angles. ...
Every developer has been there. You’re at your desk, keyboard getting sticky from coffee or dust, so you grab a microfiber cloth and start cleaning. Three seconds later you’ve somehow triggered three keyboard shortcuts, opened Slack, sent an incomplete message to a colleague, and switched to a random Neovim buffer. All because your fingers brushed the keys while wiping. The problem is stupid. The solution was obvious. So I built it. ...