Cloudbox

Generate synthetic cloud computers for training and evaluating long-horizon agents on Cloudflare.

Cloudbox turns a persona into a realistic work computer: profile, filesystem, artifacts, collaborators, multi-day simulation, and retrospective lessons. The demo opens on a completed seeded computer, so you can understand and use the repo in under seven minutes.

What you get

One public Worker URL	UI + API + static assets
Synthetic computer pipeline	persona → profile → filesystem → artifacts → collaborators → simulation → retrospective
Cloudflare primitives	Worker, D1, R2, Queue, Workers AI-ready model path
Downloadable artifacts	DOCX/XLSX/PPTX/PDF-shaped portable v1 artifacts
Full-paper mode	20-workday simulation path, queueable for longer runs

60-second architecture

Browser
  │
  ▼
Cloudbox Worker ───────────────┐
  │                            │
  ├─ /api/demo                 │ seeded complete computer
  ├─ /api/generate             │ deterministic + Workers AI-ready generation
  ├─ /api/artifacts/:id        │ downloadable artifacts
  └─ /api/runs                 │ enqueue longer full-paper work
       │                       │
       ├─ D1                   │ computers, runs, events
       ├─ R2                   │ generated artifacts + exports
       └─ Queue                │ long-running simulations

Run locally

bun install
bun alchemy dev

Open the local URL, then click through Filesystem, Artifacts, Collaborators, Simulation Log, and Retrospective.

Repo map

cloudbox/
├── alchemy.run.ts                  # Cloudflare infra in one TypeScript file
├── apps/web/                       # Worker API + static UI
├── packages/synthetic-computer/    # paper pipeline domain model
├── packages/artifacts/             # artifact exporters
├── packages/evals/                 # retrospective/rubric logic
├── docs/                           # public-facing docs site
└── migrations/                     # D1 schema

API

• GET /api/demo returns the seeded complete Cloudbox.
• POST /api/generate creates a new Cloudbox from { "text": "...", "mode": "demo" | "short" | "full-paper" }.
• POST /api/runs queues a longer run.
• GET /api/artifacts/:id downloads a generated artifact.
• GET /api/export downloads the manifest for the seeded Cloudbox.

Stack

• Workers for the public app and API.
• D1 for generated computers, run state, and event history.
• R2 for artifact objects and exports.
• Queues for long-running simulation work.
• Workers AI as the default generation path.
• Alchemy for Cloudflare infrastructure as TypeScript.

What works today

• Seeded full-paper Cloudbox opens immediately.
• Persona-to-simulation pipeline runs locally and in the Worker.
• Artifact downloads are generated and stored in R2 when available.
• D1 persistence is active when the Worker has the binding.
• Queue consumer completes queued run records.

License

MIT