The rest of my setup
The tools I did not write but use every day, and a few popular ones I tried and set aside.
Omarchy, the operating system
Omarchy is DHH's Linux distribution: Arch Linux with the Hyprland tiling window manager, configured and themed out of the box. Its site calls it "the malleable OS for the age of agents". It is my daily desktop, and I think it is the best choice for developers today.
- Keyboard first: every window, workspace and app is a shortcut away, which suits a day spent in terminals
- Arch underneath, so the newest compilers, runtimes and agent CLIs are one package away
- Everything is plain configuration files, which is exactly what a coding agent can read and change for you
- ai-usagebar and my games launcher have native Omarchy integrations
I believe in it enough to have donated to the Omacom Foundation, the nonprofit that funds Omarchy and the open source projects it is built on.
tmux, so nothing dies by accident
tmux is a terminal multiplexer: it runs your shells inside a session that lives apart from the terminal window, split into windows and panes. I use it mostly for one reason. If I close the terminal by accident, the agents keep running, and one command brings everything back.
tmux new -s work # start a named session# ...terminal closed by accident...tmux attach -t work # everything is still thereEach project gets a tmux window with its agent, and that is all the orchestration I need.
tmux on GitHubMy favorite models
I test every major model in my benchmark, but day to day the choice is simple. If you pay for an Anthropic or OpenAI subscription on the Pro plan, you are well served.
The daily driver
The latest Opus or the latest GPT Sol, at medium effort. That covers about 90% of what I do.
In the rotation
Kimi K3 and GLM 5.3, on their own subscription plans. When Claude or GPT hits its quota, I switch to one of them, and ai-memory brings the context along.
Tried and set aside: Herdr
Herdr is an agent multiplexer written in Rust: tmux-style panes and sessions, plus a sidebar that knows which pane holds a coding agent and whether it is working, blocked or done.
I tried it. It is well made and plenty of people like it. I did not find a use for it in my own routine, because one tmux window per project already tells me where each agent is, and yours may be different.
herdr.devTried and set aside: agent orchestrators
Orchestrators run several agents in parallel, each on its own git worktree, and give you one place to review what they did. The best known today:
Conductor
A macOS app: start several Claude Code or Codex agents, each in an isolated worktree, and watch their diffs from one window.
Claude Squad
A terminal app that gives each agent (Claude Code, Codex, Gemini, Aider) its own tmux session and git worktree.
They solve a real problem for some people. My bottleneck has always been how much each agent remembers, and running more of them in parallel does nothing for that.
What matters most: memory
An agent that starts every session from zero repeats the same mistakes, whichever orchestrator sits on top. Without memory, each session is an intern with amnesia. That is why I built ai-memory, and why every agent I run starts with ai-memory run.
Everything here is open source
The tools, the skills and the benchmark are public repositories. Fork them and build the version that fits the way you work.