Computers for AI agents. Each user gets a persistent cloud computer — a Linux desktop in a microVM — with a resident agent that any model can drive, plus a native macOS client to watch it, talk to it, and take over. Local coding agents can use the box through MCP.
macOS app / CLI / local agents (MCP)
│
▼
cocoonbox-api ──► control plane + edge: users, boxes, sharing,
│ Connect (third-party accounts), vault
▼
sandboxd (cocoonstack/sandbox) ──► warm pools, snapshot, clone
│
▼
cocoon microVM ── the box
├─ Xorg + XFCE desktop
├─ cocoonbox-agent ──► agent loop + MCP tools
└─ toolhost ─────────► cua-driver (loomhand/cua),
desk AT-SPI runtime
box screen ──► RFB ──► macOS app (loomhand/swift-rfb)
model calls ──► gw (cocoonstack/gateway)
cocoonbox-api in Go, the macOS app
CocoonBox.app and the cocoonbox CLI in Swift, guest image provisioning, and
the gRPC contracts.cua-driver, the screen executor in the box, with the
fixes sent upstream
(repo).loomhand builds on cocoonstack. The infrastructure stays there; loomhand holds the product and its client-side libraries.
sandboxd warm
pools, snapshot and clone; boxes are claimed from it.gw) that model calls go through.Everything lives under github.com/loomhand.