Platform
A room is a machine, not a prompt.
Every box is a harness: a fixed system prompt, a fixed tool set, a fixed output contract. When you open a room we start an agent session inside an isolated sandbox, stream its events to your browser, and throw the whole thing away when you check out.
Session lifecycle
Five states, one direction.
01
Created
02
Environment ready
03
Working
04
Idle / Waiting for you
05
Deleted on checkout
01
Create
POST /api/rooms with your brief and files. We create an agent session with the box harness and an environment image.
02
Provision
A Linux sandbox boots with the box tools mounted and your uploads in /workspace/inputs.
03
Stream
Events are relayed over SSE: messages, command executions, searches, artefacts. The UI is a pure reduction of that stream.
04
Steer
Your follow-ups are injected into the running turn. Stop cancels the turn without closing the room.
05
Check out
The turn is cancelled, the session is deleted, the sandbox is destroyed. Downloads stay live until you close the tab.
The harness
Why a box beats a blank prompt.
Fixed scope
The agent only knows how to do one job. It cannot wander, and it cannot be talked out of the output contract.
Real execution
It writes code and runs it in the sandbox, then reads the output. A broken spreadsheet is caught in the room, not by you.
Named deliverables
Every box declares its formats up front. The run is not finished until those files exist on disk.
Data handling
What we keep: nothing.
Rooms are deleted on checkout
Checking out cancels the active turn and deletes the session. The sandbox filesystem, including your uploads, goes with it.
Artefacts are yours until you leave
Deliverables are downloadable for as long as the room is open, and the links keep working until you close the tab.
Uploads stay in the room
Files you attach are written to /workspace/inputs inside the sandbox and are never copied elsewhere.
No training on your work
Briefs, transcripts and deliverables are not used to train models.