I turned my AI coding agents into an "office" so anyone can run them

I'm the maker of AI Office, a free and open source (MIT) desktop app for macOS (Apple Silicon) and Windows.

Why I built it

AI coding agents like Claude Code are powerful, but they live in terminal windows. If you are not a developer, a terminal is a wall. Even if you are one, running several agents at once gets messy: which window is working, which one is waiting for my OK, and what did it actually change?

So I moved the whole thing into an office.

• AI agents are staff. You hire Claude or GPT staff, give them work in plain chat, and they report back when they're done.
• You can see them work. Each agent sits at a desk in an office view. At a glance you see who is working, who is resting, and who is waiting for you.
• Risky stuff needs your approval. File changes and shell commands land in an approval inbox first. Nothing risky runs until you say yes, and you can undo a change.
• Big jobs go to a team lead. The lead plans the work, splits it up, and asks you to confirm before anything starts.

You don't need to know commands. "Hire staff, give work, approve" is enough. (Your company also levels up as staff work, and the office grows. A small thing, but it makes it fun.)

What you need

Claude Code or ChatGPT (Codex) already signed in on your computer. No API key. Usage is billed to that account, so running many staff at once uses your limits faster.

Everything stays on your computer. There are no analytics. The only network calls are update checks and downloads.

Not great yet

• Not code-signed, so macOS shows an "unidentified developer" warning the first time (System Settings > Privacy & Security > Open Anyway).
• Apple Silicon only on Mac, no Intel Mac or Linux.
• On Windows, notifications from outside staff, question handling, and screen capture are missing.

Try it / feedback

• Site and demo: https://engchoon.github.io/ai-office/
• Source: https://github.com/EngChooN/ai-office

I would love to hear where setup got stuck, and where the approval flow felt annoying. Especially if you are not a developer: was the "office" idea easier to understand than a terminal?