Why I moved my homelab jobs from cron to systemd timers
Cron silently skips jobs missed during downtime. Systemd timers vs cron: Persistent=true, OnCalendar scheduling, journald logs, and real service dependencies.
2026
November
October
MCP tool poisoning: approving a server once isn't approving it forever
MCP tool poisoning hides inside a server's tool descriptions. How the rug pull variant works, and a hash check that catches when a trusted tool quietly changes.
PXE booting Omarchy without taking over your router's DHCP
How to PXE boot an Arch based installer like Omarchy on a home LAN with dnsmasq proxy DHCP, GRUB netboot and nginx, without replacing your router's DHCP.
Running parallel Claude Code sessions with git worktrees
How git worktrees let multiple Claude Code sessions work on the same repo at once, and the gotchas: shared stash stacks, detached HEAD, and branch locks.
Why my Claude Code status line is 1,500 lines of bash
Building a custom Claude Code status line in bash: the stdin JSON contract, caching slow checks, Docker health over SSH, and bash 3.2 compatibility quirks.
September
Flagging ad and tracker domains with an LLM before they hit my blocklist
A dry-run daily job that asks an LLM whether a domain seen by AdGuard Home or privoxy looks like an ad or tracker, before any blocklist ever gets touched.
Auditing my own homelab backup script
A homelab backup strategy audit: what a self-hosted backup script actually protects, what it leaves in plaintext, and why a second LAN box is not offsite.
Building a Claude Code agent for a self-hosted network controller
How I wrapped a self-hosted Omada network controller's undocumented REST API in a shell script and turned it into a Claude Code agent for daily network checks.
Self-hosting a password manager with Vaultwarden and a Cloudflare Tunnel
How I run Vaultwarden, a self-hosted password manager compatible with Bitwarden clients, on my home server behind a Cloudflare Tunnel with no open ports.
Switching this blog's automation from Managed Agents to scheduled routines
Why I dropped Anthropic's Managed Agents product for this blog's post pipeline and moved to a Claude Code scheduled cloud routine on my existing Team plan.
Ideal state artifacts: tests for your AI agent config
AI coding agent setups drift like code but have no test suite. Ideal State Artifact files fix that by turning tacit expectations into checkable claims.
August
Ratusz AI: turning a hobby project into a product for town halls
A weekend search tool for one town's council resolutions became a multi-tenant SaaS in a week: what it does, how it cites sources, and what's still unproven.
What it took to send one email from a locked-down Google Workspace account
Getting kontakt@ to send mail from a locked-down Workspace account, with App Passwords and Cloud Platform both switched off, took three dead ends and one working fix.
AI news digest, week 32
Claude Opus 5 matched flagship intelligence at last gen's price, Meta undercut on agent pricing, and the EU AI Act's transparency rules went live. Weekly roundup.
July
How this blog runs itself
A static blog that writes, builds, and deploys posts with one git push. Hugo, Docker, GitHub Actions on a self-hosted ARM runner.
Office EU: a European alternative to Microsoft 365
Office EU launched in March 2026 as a fully European cloud productivity suite. Here is what it actually is, what it runs on, and whether it is worth switching.
Claude Code /btw: asking questions while Claude is still working
The /btw command in Claude Code lets you ask side questions in a floating overlay while the main task runs. Here's how it fits into a real workflow.
Exposing a local RAG server to claude.ai without opening a port
How I wired a self-hosted RAG server into claude.ai as a custom connector using Cloudflare Access, service tokens, and an OAuth portal, no port forwarding.
June
gws CLI: the missing piece of Claude Code
One CLI for all Google Workspace APIs, JSON output, bash-friendly. Here's how I replaced two MCP servers with gws and what it means for AI agent workflows.
Self-hosting Trigger.dev v4 on an ARM64 board
Running Trigger.dev v4 on a 7.4 GB ARM64 board alongside a dozen other services. What works, what eats memory, and how to survive ClickHouse.
Building a memory compiler for Claude Code
How I built a nightly pipeline that compiles raw Claude Code session logs into structured knowledge articles, stored in pgvector and mirrored to Obsidian.
Skills and agents: building reusable workflows in Claude Code
How Claude Code skills and agents turn repetitive work into reusable workflows. 19 agents, automatic skill loading, and real examples.
May
Anthropic open-sourced their skills repo
Anthropic's skills repo is now public on GitHub. What it contains, how skill-creator works, and what it means for Claude Code users.
Claude Code hooks: building a safety layer for your AI assistant
How I use Claude Code hooks to block dangerous commands, scan for secrets, inject context, and get Telegram notifications when something needs attention.
The Toggl commands that barely count as commands
Eight slash commands for Toggl time tracking: some trivially simple, one genuinely useful, and what that says about automation ROI.
The devops commands that replaced my runbooks
Eleven slash commands for deployments, diagnostics, and system health — and how they compare to the runbook documents they replaced.
April
Writing custom commands for Claude Code
How Claude Code slash commands work: file structure, $ARGUMENTS, frontmatter, namespacing, and the patterns that make them useful.
What is MCP and how to set it up in Claude Code
MCP is an open protocol that lets AI assistants talk to external tools and data sources. Here is how it works and how to configure it in Claude Code.
Hello world from a managed agent
A test post published by a managed AI agent running inside an automated pipeline. If you're reading this, the plumbing works.
The 120-file Claude Code developer workbench
48 commands, 18 agents, 22 rules, 11 hooks. How I turned a CLI coding assistant into a full development environment for DevOps and automation.
March
When your Claude Code setup audits itself
Running 5 parallel research agents to audit my Claude Code workbench: what the latest AI engineering research says, what I already had, and what was actually new.
Replacing Obsidian with pgvector: docs that live in the database
How I deleted 120 Obsidian markdown files and moved all internal docs into a local pgvector RAG system, with a rag_save MCP tool for direct ingestion.
Building a local RAG system as persistent memory for Claude Code
How I built a local multimodal RAG with pgvector, Gemini embeddings, and an MCP server so Claude Code can search past sessions and internal docs on demand.
AI news digest, week 11
GPT-5.4 landed, nine models dropped in four weeks, and AI productivity benchmarks finally have real data. Weekly roundup of what actually matters.
AI news digest, week 10
Claude Code ecosystem growing fast, MCP security under scrutiny, multi-agent orchestration going mainstream. Weekly roundup of what caught my eye.