All posts, newest first.
One Mac mini runs several AI agents across four repos. Orca is the layer that keeps that sane: every task becomes a git worktree with its own terminals, browser tabs and agent sessions, a CLI lets agents drive all of it programmatically, and a pairing code turns my phone into a second screen on the same runtime.
One container routes all my agents' LLM traffic. Getting it healthy took three fixes: container DNS that sidesteps poisoned resolver records via Clash fake-IPs, admission control tuned for agent fan-out, and a rate-limit "timeout" that was secretly killing long-reasoning models mid-flight.
I needed a browser-based web client to reach my home machine from a work PC that locks down software installs. RustDesk, RayLink, and ToDesk all gate the web client behind a paid plan, so I self-hosted HopToDesk instead, and wrote the missing web-client relay bridge along the way. Notes on what worked, what surprised me, and the one TLS gotcha that ate an afternoon.
Moving oMLX's model library to an external SSD on a Mac mini, then letting Kopia take hourly deduplicated snapshots to a second drive — so 36 GB of local models restore in minutes instead of hours of re-downloading.
Wiring Claude Code to a fully local memory stack: oMLX serves the LLM and embeddings, pgvector stores the vectors, and the open-source mem0 MCP server glues it together over SSE.
A pragmatic walkthrough: install Hugo, scaffold a site, write three posts, and build a static bundle you can open in any browser.
How a small set of type and spacing tokens can hold up a whole site without ever needing a redesign.
Restraint, proximity, and the case for letting a 1px border do the work of a 24px blur.
A 20-minute primer on getting a static site from blank directory to something worth iterating on.