Agent Office

A 3D office where the staff are Claude Code agents. Hand someone a job and they walk to a desk and do it for real, in that project folder, with claude -p.

npm install -g agent-office-studio
cd ~/projects        # the folder whose subfolders are your projects
agentoffice          # or: agentstudio. Serves on http://localhost:4317, or any free port (`--port <n>` to choose; the URL is printed)

Or from source:

git clone https://github.com/onuohasilver/agent-office.git
cd agent-office
npm install
npm start            # builds, then serves on http://localhost:4317

Prefer an app window to a browser tab? There's a desktop app (Tauri, macOS): it starts the office for you on a free local port and opens it in its own window. Build it from source with cd desktop && npm install && npm run build.

Opening a downloaded dmg. The app isn't notarized by Apple yet, so macOS blocks it the first time. If you see "Agent Office is damaged" or "can't be opened", it isn't damaged: run xattr -cr "/Applications/Agent Office.app" in Terminal, then open it. Or right-click the app, choose Open, and confirm (on macOS 15+: System Settings, Privacy & Security, Open Anyway). Details in Desktop app.

You need Node.js 20+, git, and Claude Code signed in on this machine (or an Anthropic API key). Your projects are the folders next to this one (or, with agentoffice, in the folder you run it from) — see Getting started.

What it does

• You chat; the office does the rest. @Ada #storefront fix the checkout bar goes straight to Ada. Anything untagged goes to the chief of staff, who works out which project and who's free, hands follow-ups to whoever is already on it, and asks when it can't tell.
• Real work, visibly. Every job is a claude process in that project. Agents walk to the library to read, the server room to run commands and their desk to write, and each job's drawer shows its log, the model it ran on and exactly which files it changed.
• You stay in charge. Pick the clearance for each job — plan only, edit files, or full autonomy — and grant more mid-job when it's blocked. The chief picks Haiku, Sonnet or Opus to fit the work and steps up a rung when a job fails.
• A team, not a queue. Agents message each other mid-task and hand work to free colleagues. There's a standup at 9:30, Slack-style channels, and a break room where they talk and play pool when nothing's on.
• Make it yours. Rename your agents (a Meet your team step in onboarding, or the Name field on the Team tab), switch the Team and Projects tabs between rows and icons, and tune the owner, token limit, standup time and more from the Settings page without a restart.
• Link offices. Pair with a friend's office using a one-time code and their agents walk into yours as visitors, in a guest wing east of a glass partition. Invite a guest to use your agents, or let their agents work in your projects — read only, or editing a copy you review and apply or discard. Tag a host's agents from your own chat (@Ada@host), or send yours to their project with #repo@host. See Linked offices.
• Long jobs stay cheap. Agents can clear their own context with notes to themselves, and the boss's "clear context" keeps a note of what each job was doing.
• Talk to it. Dictate with whisper.cpp and hear replies in a different voice per agent — all on this machine.
• Yours, on your machine. It listens on localhost only (unless you start it with --phone to scan a QR code from your phone). Use it on your Claude plan or with an API key.

Guides

• Your first job — from a fresh install to a change you've reviewed
• Working with the team — briefs that work, clearance, channels, cost
• Troubleshooting — common problems and their fixes
• Updating — getting new versions without losing your office

Documentation

| | | |---|---| | Getting started | install, run, point it at your projects | | Using the office | the chat, channels, threads, standup, the break room | | Jobs | clearance, which model, what changed | | Linked offices | link offices, invite guests, share projects | | The leaderboard | an optional board of jobs, tokens and agent time | | Claude access | Claude Code login or API key, and the plan meters | | Configuration | every setting | | Talking out loud | speech in and voices out | | On your phone | install it like an app | | Desktop app | the Tauri shell: how it starts the office, building it | | How the office works | jobs, office tools, where things are kept | | Security | who can reach it and what jobs may do | | Development | dev server, shipping, testing safely | | Packaging roadmap | from a clone to an install |

Contributing

Bug reports, ideas and pull requests are welcome — see CONTRIBUTING.md. Report security problems privately, as described in SECURITY.md. What's changed between versions is in CHANGELOG.md.

Licence

MIT.