Sandbox Skill

Isolated execution environments for AI agents and testing workflows.

Philosophy

Lambda/Functions: Fast cold start, 15-min limit, stateless, per-ms billing
App Platform Sandbox: 30s cold start (or instant with pool), unlimited duration, stateful, per-hour billing

Sweet spot: Long-running, stateful, iterative workflows where agents need to
install packages, run code, check results, modify, repeat.

Quick Decision

┌─────────────────────────────────────────────────────────────┐
│         Need isolated execution environment?                 │
└─────────────────────────────────────────────────────────────┘
                            │
              Is this for debugging an EXISTING app?
                            │
            ┌───────────────┴───────────────┐
            │                               │
           YES                              NO
            │                               │
            ▼                               ▼
   ┌─────────────────┐           ┌─────────────────┐
   │ troubleshooting │           │ Need real-time  │
   │ skill           │           │ streaming or    │
   │                 │           │ port exposure?  │
   │ Sandbox.get_    │           │                 │
   │ from_id()       │           └────────┬────────┘
   └─────────────────┘                    │
                          ┌───────────────┴───────────────┐
                          │                               │
                         YES                              NO
                          │                               │
                          ▼                               ▼
                 ┌─────────────────┐           ┌─────────────────┐
                 │ SERVICE MODE    │           │ Is low latency  │
                 │ exec_stream()   │           │ critical?       │
                 │ expose_port()   │           │                 │
                 └─────────────────┘           └────────┬────────┘
                                                        │
                                        ┌───────────────┴───────────────┐
                                        │                               │
                                       YES                              NO
                                        │                               │
                                        ▼                               ▼
                               ┌─────────────────┐           ┌─────────────────┐
                               │ HOT POOL        │           │ COLD SANDBOX    │
                               │ SandboxManager  │           │ Sandbox.create()│
                               │ ~50ms acquire   │           │ ~30s startup    │
                               └─────────────────┘           └─────────────────┘

Prerequisites

# Verify doctl is installed and authenticated
doctl auth whoami

# Install the SDK (choose one)
uv pip install do-app-sandbox
# OR
pip install do-app-sandbox

# For Spaces support (large file transfers)
pip install "do-app-sandbox[spaces]"

Requirements:

• Python 3.10.12+
• doctl CLI installed and authenticated
• DigitalOcean account with App Platform access

Quick Start: Cold Sandbox

Single sandbox creation with ~30s startup time:

from do_app_sandbox import Sandbox

# Create sandbox with Python image
sandbox = Sandbox.create(
    image="python",           # or "node"
    name="my-sandbox",
    region="nyc",
    instance_size="apps-s-1vcpu-1gb"
)

# Execute code
result = sandbox.exec("python3 -c 'import sys; print(sys.version)'")
print(result.stdout)

# File operations
sandbox.filesystem.write_file("/tmp/script.py", "print('hello')")
result = sandbox.exec("python3 /tmp/script.py")

# Clean up
sandbox.delete()

Full guide: See [cold-sandbox.md](reference/cold-sandbox.md)

Quick Start: Hot Pool

Pre-warmed sandboxes for instant acquisition:

import asyncio
from do_app_sandbox import SandboxManager, PoolConfig

async def main():
    # 1. Configure pool
    manager = SandboxManager(
        pools={"python": PoolConfig(target_ready=3)},
    )

    # 2. Start and warm up (blocks until pool is ready)
    await manager.start()
    await manager.warm_up(timeout=180)

    # 3. Acquire instantly (~500ms from pool vs 30s cold start)
    sandbox = await manager.acquire(image="python")
    result = sandbox.exec("python3 -c 'print(2+2)'")
    print(result.stdout)

    # 4. Delete when done - YOUR responsibility!
    sandbox.delete()

    # 5. Shutdown (cleans up pool, not acquired sandboxes)
    await manager.shutdown()

asyncio.run(main())

Ownership model: Once you acquire() a sandbox, you own it. Always call sandbox.delete() when done. The shutdown() only cleans up sandboxes still in the pool.

Full guide: See [hot-pool.md](reference/hot-pool.md)

Quick Reference: When to Use Sandbox

Quick Reference: Available Images

Working directory: /home/sandbox/app (with /app symlink). Ports: 8080 (user apps), 9090 (health checks).

Custom images supported — any Docker image with HTTP server capability.

Quick Reference: SDK Methods

Reference Files

• [cold-sandbox.md](reference/cold-sandbox.md) — Single sandbox lifecycle, file ops, cleanup
• [hot-pool.md](reference/hot-pool.md) — Pool management, sizing, cost optimization
• [service-mode.md](reference/service-mode.md) — Streaming, port exposure, AI agent patterns
• [use-cases.md](reference/use-cases.md) — AI agent patterns, testing patterns
• [positioning.md](reference/positioning.md) — Lambda vs Sandbox decision guide

Cost Considerations

Sandbox billing: ~$0.01-0.03/hour per container (apps-s-1vcpu-1gb)

Hot Pool trade-off:
- Pool of 5 sandboxes running 8 hours = ~$0.80-2.40/day
- Eliminates 30s cold start per request
- Worth it for interactive AI agents, not for batch jobs

Integration with Other Skills