← All editions

I built a personal AI. Here's the stack.

Last week I asked an AI to scan my MCP server configs for security vulnerabilities, build a CLI tool to automate the checks, add baseline diffing, and publish it to npm. It did all of it across a few sessions. No copy-pasting from Stack Overflow. No scaffolding by hand. Conversations that produce working software.

That's not a demo. That's my normal workflow now.

The Stack That Makes This Work

I run what I call a Personal AI Infrastructure — Claude Code as the brain, with a skills system that gives it deep context about my work, my tools, and my preferences. Think of it like muscle memory for an AI.

Here's the actual architecture:

Claude Code (CLI)
  ├── Skills (60+ context modules)
  ├── n8n (workflow automation, self-hosted)
  ├── Obsidian (knowledge base, semantic search)
  ├── MCP Servers (browser, Docker, Obsidian)
  └── Multi-agent spawning (parallel research, builds)

The skills are the interesting bit. Each one is a markdown file that teaches the AI about a specific domain — manufacturing processes, infrastructure details, deployment patterns, even personality calibration. When I say "deploy this to the VPS", the AI already knows which server, which port range, and how to configure Traefik. No explanation needed.

The result: I don't use AI as a chatbot. I use it as a technical partner that remembers everything and can operate my infrastructure.

What this took to build: About three months of evening work. The core is just Claude Code with good context management. The skills system is markdown files in a directory. n8n handles the automation. It's not complex — it's just well-organised.

What it costs: Claude Pro subscription, a £12/month VPS, and whatever API calls the agents make. Under £50/month total.

Why The Agent Stack Exists

Most AI newsletters tell you what happened. This one shows you what to build.

I'm a practitioner. I build agents that do real work — automate research, manage infrastructure, generate reports, scan for security issues. Every edition of The Agent Stack comes from something I actually built, broke, or figured out.

Three editions per week:

• Monday: Build something. Walkthroughs with real code.
• Wednesday: Stack reviews. Tools and frameworks tested in production, not in theory.
• Friday: Signal. What happened this week that changes how you build agents.

No hype. No "revolutionary AI breakthrough" headlines. Just what works.

Quick Hits

• Claude Code now supports multi-agent teams — Spawn parallel agents that coordinate via shared task lists. I use this daily for research. Game-changer for complex builds.
• n8n raised $180M at $2.5B valuation — The workflow automation tool I use for everything. 75% of their customers now use AI tools. The agent infrastructure market is real.
• Gartner says 40% of enterprise apps will feature agents by end of 2026 — Bold claim. Based on what I'm seeing, the number for useful agents is closer to 5%. But that 5% is transformative.

One Thing to Try

Install Claude Code and create your first skill file. Make a file at ~/.claude/skills/MyProject/SKILL.md with context about a project you're working on — tech stack, deployment details, coding conventions. Next time you work on that project, the AI already knows. It's the simplest force multiplier I've found.

This is edition one. If you're reading this, you're early. That's usually the best time to be somewhere.