Agent Office – A 3‑D collaborative UI for AI coding agents

What it does – Agent Office turns a collection of LLM‑powered coding assistants (Claude Code, Codex, OpenCode, Grok, Muse, DeepSeek Harness, Pi, Cursor, …) into virtual “workers” that sit at desks inside a 3‑D office you view in a web browser. Each worker runs its own CLI, shows a live terminal on a laptop, and can be spoken to with a microphone‑enabled prompt box. Projects are organised as floors: every GitHub repository you add becomes a separate floor, complete with issue/PR cork‑boards, git worktrees, and the ability to open PRs directly from the desk.

How you use it

1. Install the agent‑office CLI (via the one‑line curl … | bash script) on a machine that has Node 20+, git, the GitHub CLI and at least one supported agent CLI installed.
2. Run agent-office. The first‑run wizard asks where to clone projects, signs you into GitHub, and lets you pick an initial repo – that repo becomes the first floor.
3. Open the generated URL in a browser; the office UI loads on 127.0.0.1 (or on a server you expose via SSH tunnel). Walk around, press E at an empty desk, and hire a coding agent of your choice.
4. Interact with the agent by typing or by holding Ctrl+Space / clicking the 🎤 icon to speak; the spoken text is transcribed client‑side and sent to the agent.
5. Agents can be instructed to create worktrees, open issues, push branches, and run gh pr create. The UI shows visual cues (red beacons, banners, sounds) when an agent needs your attention or finishes a task.
6. Teams can share an office instance by deploying it to a cloud VM (AWS, Azure, Railway, Fly.io, Dokploy, etc.) and tunnelling in with SSH or Tailscale. The server runs only on 127.0.0.1; access is mediated by the tunnel, keeping voice‑chat and screen‑share functional without public certificates.

Key features

• Multi‑agent support – pick from several LLM coding models, each with its own “reasoning effort” setting.
• Project‑per‑floor layout – each GitHub repo is a separate floor; workers get dedicated git worktrees and can open PRs in the correct repo.
• Voice‑first interaction – browser‑based speech‑to‑text, no extra software needed.
• Real‑time visual alerts – beacons, banners, and sound cues tell you which agent is waiting.
• Cross‑device UI – a full‑3D view for desktops and a /lite 2‑D view for phones or low‑power machines.
• Collaboration tools – shared whiteboard, screen‑share TV, chat, and issue/PR boards on virtual cork walls.
• Custom maps – replace the office with a castle, space‑station, or any JSON‑defined layout.
• Extensive deployment options – scripts for AWS EC2, Azure VM, Railway, Fly.io, Dokploy, Coolify, or a plain Ubuntu/Debian server; also supports Tailscale for HTTPS‑less access.
• Agent‑office MCP server – a lightweight management layer that lets agents list, hire, message, and retire other agents via the office‑workers command.

Who it’s for

• Individual developers who already use LLM coding assistants and want a visual, shared workspace.
• Small teams that need a single UI to monitor many agents across multiple repositories.
• Anyone who prefers voice interaction or wants a playful “office” metaphor for AI‑augmented development.

Limitations / status

• The project is explicitly work‑in‑progress and tailored to the author’s workflow; breaking changes are expected between releases.
• It requires the proprietary CLI for each supported agent (Claude Code, Codex, etc.) and a valid subscription or API key for those services.
• The UI runs locally or behind an SSH tunnel; exposing it directly to the internet is not the default.

Getting started

curl -fsSL https://raw.githubusercontent.com/AgentSystemLabs/agent-office/main/install.sh | bash
agent-office   # runs the wizard and opens the browser

For cloud deployment, follow the appropriate deploy/*.sh script documented in the README (AWS, Azure, Railway, Fly.io, Dokploy, etc.).