What AgentRoom is

AgentRoom lets a Vision Pro control coding-agent sessions that run on a Mac. The code, the credentials, and the agents stay on the machine you already work on. The headset is a client.

That sentence carries a design decision worth stating plainly before anything else: there is no hosted control plane and no container image. You do not upload a repository to us. We do not run your agent. There is no account to create and no server of ours in the path between you and your code.

The three pieces

The backend runs on your Mac. It registers folders, starts coding agents inside them, owns session state, and sends typed updates to the Apple clients. It is the source of truth for everything: registered workspaces, sessions, turns, messages, and the audit log.

The macOS app is the operator console. It configures and supervises the backend, stores secrets in the Keychain, and registers the folders you want agents to work in. It is deliberately not a place to send prompts from. Its threads view shows you what is happening; it does not drive the conversation.

The visionOS app is the session client. This is where you actually work. You open threads, send turns, edit files, review Git changes, and lay diagrams out in the room around you.

Why it runs on your Mac

The obvious way to build this would have been a hosted control plane. Upload the repository, run the agents in a container somewhere, stream the result to the headset. That is how most of this category works, and we did not do it.

An agent that is useful on your codebase needs your codebase. Not a copy of it, and not a snapshot from twenty minutes ago. It also needs the things around your codebase: your runner logins, your Keychain, your SSH agent, the local tools your project builds with. Sending all of that to somebody else's machine in order to get a nicer client is a bad trade, and it is a trade you cannot easily undo once you have made it.

So the backend runs where those things already are. Nothing has to move, because nothing is missing.

The cost is real and worth naming. You need a Mac, you need it awake, and you need it on the same network as the headset. AgentRoom is not something you can use from an aeroplane while your Mac sits at home. We think that is the right trade for the kind of work it is for, but it is a trade.

Why a headset

A flat monitor puts a hard limit on how much of a codebase you can hold in view. You get two or three files, and everything else is a tab you have to remember to go back to.

A coding-agent session is worse in this respect than ordinary editing, because a single turn produces several things you want to look at together. There is the conversation, the diff, the file tree, the plan the agent is working from, and often a terminal. On a laptop these compete for the same rectangle. In a headset they can each have a place, and the place persists between turns.

The spatial diagrams follow from the same idea. AgentRoom renders live SVG and Mermaid sketches, and it has editable diagrams stored as ordinary files in your workspace. They are files, not a proprietary format inside our app, so they commit alongside the code they describe.

Where to go next

• How AgentRoom is put together covers the boundaries: what the backend owns, what clients can and cannot ask for, and how a thread survives a restart.
• The coding agents AgentRoom drives covers the four built-in agents, what each needs to be set up, and what each one loads from your workspace.
• Trust and safety is the honest version: what is bounded, what ships switched off, and where the sharp edges are.
• Install AgentRoom walks through the TestFlight beta and pairing a headset to your Mac.