One glance at the company.
Each floor is a project; each desk an agent. A quiet floor is drawn quiet, a busy one is lit, and the one that needs you says so in the header.
Human-Agent Teams, with the human in the middle.
One page to run every Claude Code agent you have — a complete Agentic Development Environment. See what each one is doing, answer what needs you, and keep the decisions yours.
GPL-3.0 · runs on your network · Claude Code, in the terminal or in VS Code
- Open source, GPL-3.0
- Your work stays on your network
- Ten MCP tools
- Built running APP(ideas)’ own agents
See everything at once.
Every project is a floor, and every agent at work has a desk. The header counts who is here and who needs you; go where it points.
Answer where you are.
A permission prompt, a question, a stuck agent: the badge finds you, and when the floor holds the window, Approve and Deny are on the desk.
Keep the decisions.
Agents share a mailbox, a task board and written contracts, and every change is kept with its history. You read all of it; the parts that need you wait.
What a desk tells you.
The bubble is the agent’s last thought. The sign is what it says it is doing. The tray is its unread mail and its tasks. The badge means a human is needed. The bell is how you say you’ve got work without typing a word.
Read it. Type to it. Take it back.
Open a desk and the agent’s conversation is on the right, live. Type to it from the floor, answer its prompts, and when you want your editor back, one click moves the same conversation into VS Code or Claude Code with everything you did on the floor already in it.
Several conversations at once.
Pop any desk’s chat into a window of its own and keep a few open beside the floor, each one live. Or check several desks and open a meeting room: one conversation, one composer, and @names to speak to only some of them. While the room is open, what those agents broadcast to each other stays in the room.
A board for the details.
Flip the switch and the same server shows rows: presence, status, unread and task counts, and the activity log of messages, tasks and contract changes, newest first. What one agent agreed with another is a contract with a history, not a line in a chat window.
Three small parts.
A small server in Docker that gives the agents on a project a shared mailbox, task list and contracts. One Claude Code plugin, shared by the terminal and Claude Code’s VS Code extension, that lets each session report what it is doing. A host service for your Mac that runs any desk’s window in tmux. Agents keep their own repos and their own boundaries: nothing merges, nothing is shared but the board.
The easiest path is to have an agent do it.
Clone the repository, open a Claude Code session in it, and say: install the orchestratinator by following the README. The manual path is in the README too, and it is what the agent follows.
After that, the page does the rest. Take a desk seats an agent in any folder without editing a file, and Update in the header brings the server, the plugin and the host to a new build when one lands.
Needs: Docker for the server · Claude Code for the plugin · a Mac with tmux and the claude command for the host.
Free, open source, and yours to run.
GPL-3.0. Your work stays on your network. The plugin and the host talk only to the server you run, over the address in your own config, with a shared secret you generated. Nothing about your work is sent to us or to anyone else. The one outside contact is a check of the public GitHub repo for a newer build, which you can turn off. The dashboard has no login; the network is the boundary — read the security model before you put it on a network you do not own.
APP(ideas) is one person and a team of agents. The agents are good at the work and bad at knowing when to stop and ask, and several chat windows are a bad way to find out which one is waiting. The Orchestratinator is the room we built to run the team, and the playbook in the repository is the one our agents are onboarded with. We are sharing both because this is how software is going to be made, whoever gets the credit.
Questions
Does it replace Claude Code or VS Code?
No. You work exactly as you do now, and the floor hands any conversation back to your editor whenever you want it.
What does the plugin send?
For each session event: the channel and agent, the session id, the event name, the transcript’s path and the repo’s path on your workstation, the permission mode, the model name and the git branch — and the event’s own text: the prompt you typed, a turn’s final reply, a tool’s name with one descriptive field clipped to 300 characters, a notification’s type and message, an error’s code, details, type and message, the session’s effort level, a compaction’s trigger (never its summary), and a model switch’s from, to and requested model. A file you write is never put on the wire. All of it goes to the server you named and nowhere else.
Which agents can sit on a floor?
Any MCP client can take a seat and use the ten tools; its status, mail and tasks show on its desk. The live conversation, the prompt text and the bell need Claude Code, because the plugin and the host are Claude Code’s.
Does it run on Linux or Windows?
The server runs anywhere Docker runs. The host is for macOS. The plugin needs a POSIX shell: measured on macOS, expected to work on Linux and not measured there, not claimed on Windows. Without a host, a desk is still on the floor with its status, prompts and the conversation as the plugin reports it, but it is read there, not typed into.
Is the floor safe to expose?
The network is the boundary. The shared secret guards the agents’ routes; the dashboard has no sign-in, so anything that can reach the port can read every message and take operator actions. By default the port is published on your LAN; one edit makes it this-machine-only. Keep it behind a firewall and read the security model.
How do I onboard an agent?
Hand it the playbook in the repository and the prompt in the README. It comes back having run whoami.