The whole setup for running coding agents side by side, out of one box. You see which one needs you, and you never reach for the mouse.
macOS or Linux, paste into a terminal
git clone https://github.com/ZyxWorks/agent-office.git ~/agent-office && ~/agent-office/bin/office tools && ~/agent-office/install.shthen, every morning
office onoffice tools lists every tool you lack and asks before it installs
one, each at a pinned version from its own maker. install.sh shows every file it
will write and backs up what it replaces. Neither starts anything. You bring git, Node 22.19 or
newer, and a coding agent you signed in to yourself, such as Claude Code or Codex. One thing
is left to you before the first office on: clone
Firstmate and put its path in the config.
MIT licensed, and all of it is on GitHub before you run any of it: the installer is one short script you can read first, and so is the list of everything it installs. Getting started walks you through the first day.
Running three or four coding agents at once is normal now. The tools for it exist, and they are good: herdr runs the terminal and knows each agent's state, Firstmate supervises a fleet of workers, treehouse hands each one its own git worktree. But each is its own install, its own config and its own update, and the wiring between them lived in one person's dotfiles. A new machine meant doing it all again by hand. A new colleague meant "how did you set that up".
Agent Office is that wiring, packaged and tested. It does not rewrite any of those tools. Each one stays with its own maker.
Several agents you can keep track of is a different tool from several agents in several windows.
Every agent has one line in the sidebar. herdr knows whether it is working, idle, done or blocked. The Agent Office meter adds the two numbers that decide what you do next: how full its context is, and how long it has been quiet.
claude · 412k▲ · 47m · idle| you see | it means |
|---|---|
412k | tokens in this agent's context right now |
▲ amber | past 200k, or past half of a smaller window |
▲▲ red | past 600k, or past three quarters |
47m | time since it last wrote to its transcript |
| a push on your phone | an agent is blocked. Off until you turn it on |
The numbers come from the agent's own transcript: Claude Code, Codex, and Claude Code on a local Ollama model. An agent the meter cannot identify stays blank. It never guesses. Both limits are yours to set in the config.
One key walks the whole office. Shift+← and Shift+→ go to the pane on that side, and from the last pane to the next tab. Shift+↑ and Shift+↓ do the same, and from the last pane to the next space. It stops at the ends, so you can hold it.
Everything else is Ctrl+Space, then one letter. A small letter is for the space, the same letter with Shift is for the tab.
| you want to | after Ctrl+Space |
|---|---|
| make a new space, or a new tab | n N |
| close one | c C |
| rename one | r R |
| split the pane to the right, or below | v - |
| go to any agent by name | g |
| go to the agent that asked for you | o |
| leave, with everything still running | d |
office keys prints every key on one screen. A probe checks that
sheet against herdr itself on every push, so it cannot drift from what the keys really do.
| you want to | you type |
|---|---|
| open the office and its Firstmate | office on |
| talk to Firstmate | fm |
| walk away and leave it all running | office break |
| see what is running | office status |
| see what is installed, and whether it is a tested version | office doctor |
| move every tool to the version this release names | office update |
| go home | office off |
You tell Firstmate what you want done. It starts the workers, each in its own worktree, and reports back. You watch the sidebar.
office off stops the office it started, after it shows you what will stop. It never stops every claude on the machine, and it never deletes a worktree.Every command on this page has a probe that runs it in a throwaway home, on macOS and Linux, on every push. The source is public, and so is every claim on this page.
ZyxWorks is a product studio and forward deployed engineering practice. Agent Office is one of the tools we built for our own work and gave away. The paid work is putting the same engineering inside a company.