Repository
Pull requests become tasks: draft = in progress, open = in review, merged = done. claude/, codex/ and copilot/ branches sit at that agent's desk.
GitHub token (optional)
Only needed to see CI on the monitors and to post an "ask" comment. Use a fine-grained token for just that repository with Pull requests: read and write and Checks: read. It stays in this tab's memory: never saved, and sent only to api.github.com. Close the tab to forget it.
Voice
Controls
| Control | What it does |
|---|---|
| Hold grip or A/X | Push to talk. Let go and it acts on what you said. With hand tracking, pinch the Talk button on the panel instead. |
| trigger (or pinch) | Select whatever the ray points at: a sticky note, a monitor, a panel button. |
| Left stick | Walk, in the direction you're looking. |
| Right stick | Snap turn left or right. |
| Preview | Drag to look, WASD to walk, click to select, hold Space to talk. |
Things to say
| Say | What happens |
|---|---|
go to the board · go to Claude's desk · recenter | Move to a spot in the room. |
refresh · check CI | Pull the repository's pull requests again (and, with a token, their check runs). |
select 3 | Highlight task 3 so the panel's Start, Review and Done buttons act on it. |
start 2 · move login to review · finish 4 | Move a task. The same commands as the 2D room, which shares this board. |
fix the flaky upload test for Builder | Anything that isn't a command is a new task. |
ask Claude on 12 to add a regression test | Drafts a pull-request comment that starts @claude, for agents that answer mentions (the Claude Code GitHub action, Copilot, Codex). Nothing is posted until you say confirm or press Send; cancel drops it. Needs a token. |
help | Reads the list back. |
Questions
Which headsets work?
Any browser with WebXR's immersive-vr mode: Meta Quest Browser is the main target. The Enter VR button only lights up where the browser offers it. Everywhere else the preview above shows the same room on a screen. This page was built and tested in a desktop browser without a headset, so expect rough edges on a real device and please report them.
Does my voice go anywhere?
Where the browser has built-in speech recognition (Chrome, Edge, Android), that is used, and some browsers, Chrome among them, send the audio to the browser maker's speech service. Where it doesn't (Quest Browser, Firefox), the page runs a small Whisper model (tiny, English) on your own device: the model is downloaded once from Hugging Face via the jsDelivr CDN and your audio never leaves the headset. Tick Always use on-device speech to use it everywhere. Recognition is push-to-talk only: the microphone is open only while you hold the button.
Can it really talk to my agents?
It talks to them where they already listen: on GitHub. "Ask Claude on 12 to …" posts a comment on pull request 12 that starts @claude, which the Claude Code GitHub action (and Copilot or Codex for @copilot / @codex) picks up if the repository has it installed. The comment is posted as you, with your token, only after you confirm, and says it was sent from Agent Workroom. Agents running in a terminal or a cloud session can't be reached from a web page; for those, use Copy brief on the 2D page.
Where are my tasks?
In this browser's local storage, shared with the 2D room on the same site. Nothing is uploaded to a server of ours.