Agent Office turns Claude Code, Codex and OpenCode into workers at desks in a shared 3D office: Reveneau

Why it mattersA team running several coding agents at once now has one shared view of what each one is doing, who needs a decision, and which pull request belongs to which issue, instead of a wall of terminal tabs on one person's laptop.

Hiring one coding agent is easy. Watching four of them at once, on three different repos, without losing track of which one is stuck on a decision, is a different job. Agent Office is a new open-source project that gives the whole team one place to do it.

The repository at AgentSystemLabs/agent-office was created on 26 September 2026 and reached 380 stars, 81 forks and 20 open issues in the three days that followed, per the GitHub API on 29 September. It is written in TypeScript, licensed MIT, and installs with one shell command on macOS, Linux or Windows.

What the office looks like

Each GitHub repository the team connects becomes a floor of a cartoon office building. Empty desks sit around the floor. Walk up to one in your browser, press E, and pick one of three coding agent CLIs to sit at that desk: Anthropic's Claude Code, OpenAI's Codex, or OpenCode. The worker's live terminal shows on the laptop in front of it, and anyone in the office can open that terminal and type. When a worker finishes or needs input, it jumps up and dings, and one key press takes you to the one that has waited longest.

GitHub issues and pull requests hang on cork boards on the walls of each floor. A team member can hand an issue to a worker, queue tasks, give a worker its own git worktree, and open a pull request with one key. The README says one task can span several projects: the worker gets a worktree of each, and a pull request in each that links the others.

What connects the pieces

Everything runs on one machine. The requirements list is Node.js 20 or newer, git, the GitHub CLI signed in for the boards, and at least one of the three coding agent CLIs signed in with the user's own account. Voice, chat and screen sharing on a lounge TV are built in for the people in the office. The default is a solo setup on the machine that started it, on 127.0.0.1, so nobody else can reach it.

To share the office with a colleague, the project ships one-script deployments to AWS EC2, Azure and Railway, each launching a small Ubuntu virtual machine, running the office under systemd, and putting it behind an SSH tunnel or a Tailscale hostname. The AWS script launches a t3.xlarge (four virtual CPUs, 16 gigabytes of memory) and states the office listens on 127.0.0.1:4600 on the machine so there is no public URL to protect.

The state of the pitch

The README is honest that the project is early. The maintainer marks it a work in progress built for one person's workflow, warns that breaking changes should be expected between releases, and invites anyone whose needs are close to fork it and bend it. The traction is real (380 stars in seven days, 81 forks, verified 29 September through the GitHub API) and the project does not claim adoption numbers or performance measurements. Nothing was independently tested here.

For a small team that already has more than one coding agent open at a time, on more than one repository, the office is a place to see them side by side rather than to add capability. The trade is that the whole team can look at what any worker is doing, at the cost of one machine somebody has to run.

Source

GitHub repository, primary source: https://github.com/AgentSystemLabs/agent-office
GitHub REST API for star, fork and issue counts, queried 29 September 2026: https://api.github.com/repos/AgentSystemLabs/agent-office

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.