A 3D office where real Claude Code sessions work as employees

Developer brstic launched Orchestrator, an open-source 3D virtual office environment designed to manage autonomous AI coding agents as in-person desk workers, detailed in github.com's report. The platform runs on a persistent server where every worker seated at a desk is an active Claude Code process executing inside an isolated tmux session and dedicated git worktree.

This report covers the architecture, game-style interface, task orchestration pipeline, account rotation systems, and setup requirements of Orchestrator. The release serves engineering teams, tech leads, and autonomous agent operators seeking visible multi-agent oversight without tabbing through dozens of unmonitored terminal windows.

01What shipped in A 3D office where real Claude Code sessions

Orchestrator provides a persistent multiplayer 3D floor plan navigable using standard W A S D controls. Users interact directly with agents by pressing F to open conversation dialogs or pressing Space to sit at an agent's terminal. Agents physically act out their real terminal operations across the floor, holding their chin in hand while thinking, typing while running code, and raising a physical hand when blocked by a pending permission prompt or decision.

Beyond the workspace simulation, the platform packages complete multiplayer collaboration features over TLS. Colleagues can join the shared office through a dedicated desktop app or a web browser, bringing their local machines onto the office floor as dedicated private rooms. The environment also includes an interactive two-player arcade cabinet running on an emulated Mega Drive, an office break room featuring internet radio, an in-office browser television, and a rooftop terrace where avatars can parachute into an endless driving city.

02What improved in A 3D office where real Claude Code sessions

The system replaces manual terminal-tabbing with deterministic state tracking tied directly to machine events, hooks, and tool outputs. When tasks encounter permission prompts or questions, requests remain visibly held open across full system restarts until human operators respond. The operational differences between managing individual terminal sessions and deploying workers inside Orchestrator are measured across key operational metrics.

Autonomous delivery under this framework was demonstrated when a single simulated office worker built twenty-one distinct interactive 3D website previews for commercial businesses in a single day. These builds spanned bespoke interactive models including agricultural greenhouses, solar calculators, functional tractor X-rays, and moving mechanical assembly lines while the rest of the agent roster worked uninterrupted on parallel projects.

03What you gain from A 3D office where real Claude Code sessions

Engineering organizations gain strict verification barriers where task completion is enforced by runtime checks rather than agent claims. Every assigned task automatically generates an isolated branch and git worktree on the host machine. Completion requires that local test suites execute and pass on the machine, an agent colleague conducts a review, the human owner approves the diff, and the branch merges cleanly without merge conflicts.

Task management is augmented by an autonomous foreman supervisor that walks the office floor to answer routine operational questions, freeing human engineers to focus on architectural direction. Teams can issue persistent /goal instructions that keep agents executing across multiple turns until target test suites pass completely. Workflows also incorporate repeated timed routines, named secret management through an integrated vault, and long-term memory persistence across sessions via MemPalace.

Multi-account pooling ensures development does not halt when individual API limits are reached. Machines can register multiple accounts alongside Z.AI GLM plans, allowing sessions to automatically switch to alternative accounts or cascade from Claude 3.7 Fable to Opus, Sonnet, or Haiku. Inactive sessions sleep automatically after thirty minutes of idle time and resume immediately when pinged by a task, message, or the operator whip hotkey.

04How to get A 3D office where real Claude Code sessions

Deploying Orchestrator requires Node.js version 22.18 or higher, tmux, git, pnpm, and an authenticated Claude Code CLI installation on the host machine. Operators install dependencies and compile the workspace bundles using pnpm:

pnpm install
pnpm build

To run the application locally with an integrated server and local agent host, start the runtime in browser mode or launch the standalone desktop client:

pnpm start --open

pnpm desktop

Within the application interface, navigate to Menu, select Team, and choose Hire to configure worker profiles. Operators can specify custom names, assign roles, define host folder directories, and select default models and effort tiers. In-app switches support targeted model selection or fallback assignments:

/model claude-fable-5-1

claude --model fable

05What to watch after A 3D office where real Claude Code sessions

Teams implementing the platform must monitor machine resource allocation as concurrent tmux sessions and worktrees scale. While the office front door allows external visitors to knock and registered members to sign in directly, multi-tenant agent execution requires careful isolation between shared server resources and individual contributor laptop rooms. Organizations running remote teams should host the central office instance on a persistent Linux server to maintain uninterrupted background routines and continuous test verification.

Operational tracking must also monitor account failovers when running complex goal loops. Because the office counts active execution time to assign monthly employee rankings, teams should audit background command indicators to prevent long-running test servers or compilation builds from holding active session state. Future releases are slated to expand external attachment handling and expand browser automation tooling used by agents during review cycles.