cloudflare-sandbox

Cloudflare Sandboxes SDK

Status: Production Ready (Open Beta) Last Updated: 2025-10-29 Dependencies: cloudflare-worker-base, cloudflare-durable-objects (recommended for understanding) Latest Versions: @cloudflare/sandbox@0.4.12, Docker image: cloudflare/sandbox:0.4.12

Quick Start (15 Minutes)

1. Install SDK and Setup Wrangler

npm install @cloudflare/sandbox@latest

wrangler.jsonc:

{
  "name": "my-sandbox-worker",
  "main": "src/index.ts",
  "compatibility_flags": ["nodejs_compat"],
  "containers": [{
    "class_name": "Sandbox",
    "image": "cloudflare/sandbox:0.4.12",
    "instance_type": "lite"
  }],
  "durable_objects": {
    "bindings": [{
      "class_name": "Sandbox",
      "name": "Sandbox"
    }]
  },
  "migrations": [{
    "tag": "v1",
    "new_sqlite_classes": ["Sandbox"]
  }]
}

Why this matters:

• nodejs_compat enables Node.js APIs required by SDK
• containers defines the Ubuntu container image
• durable_objects binding enables persistent routing
• migrations registers the Sandbox class

2. Create Your First Sandbox Worker

import { getSandbox, type Sandbox } from '@cloudflare/sandbox';
export { Sandbox } from '@cloudflare/sandbox';

type Env = {
  Sandbox: DurableObjectNamespace<Sandbox>;
};

export default {
  async fetch(request: Request, env: Env): Promise<Response> {
    // Get sandbox instance (creates if doesn't exist)
    const sandbox = getSandbox(env.Sandbox, 'my-first-sandbox');

    // Execute Python code
    const result = await sandbox.exec('python3 -c "print(2 + 2)"');

    return Response.json({
      output: result.stdout,
      success: result.success,
      exitCode: result.exitCode
    });
  }
};

CRITICAL:

• MUST export { Sandbox } from @cloudflare/sandbox in your Worker
• Sandbox ID determines routing (same ID = same container)
• First request creates container (~2-3 min cold start)
• Subsequent requests are fast (<1s)

3. Deploy and Test

npm run deploy
curl https://your-worker.workers.dev

Expected output:

{
  "output": "4\n",
  "success": true,
  "exitCode": 0
}

Architecture (Understanding the 3-Layer Model)

How Sandboxes Work

┌─────────────────────────────────────────┐
│  Your Worker (Layer 1)                  │
│  - Handles HTTP requests                │
│  - Calls getSandbox()                   │
│  - Uses sandbox.exec(), writeFile(), etc│
└──────────────┬──────────────────────────┘
               │ RPC via Durable Object
┌──────────────▼──────────────────────────┐
│  Durable Object (Layer 2)               │
│  - Routes by sandbox ID                 │
│  - Maintains persistent identity        │
│  - Geographic stickiness                │
└──────────────┬──────────────────────────┘
               │ Container API
┌──────────────▼──────────────────────────┐
│  Ubuntu Container (Layer 3)             │
│  - Full Linux environment               │
│  - Python 3.11, Node 20, Git, etc.      │
│  - Filesystem: /workspace, /tmp, /home  │
│  - Process isolation (VM-based)         │
└─────────────────────────────────────────┘

Key Insight: Workers handle API logic (fast), Durable Objects route requests (persistent identity), Containers execute code (full capabilities).

Critical Container Lifecycle (Most Important Section!)

Container States

┌─────────┐  First request  ┌────────┐  ~10 min idle  ┌──────┐
│ Not     │ ───────────────>│ Active │ ─────────────> │ Idle │
│ Created │                 │        │                │      │
└─────────┘                 └───┬────┘                └──┬───┘
                                │ ^                      │
                                │ │ New request          │
                                │ └──────────────────────┘
                                │                         │
                                ▼                         ▼
                            Files persist          ALL FILES DELETED
                            Processes run          ALL PROCESSES KILLED
                            State maintained       ALL STATE RESET

The #1 Gotcha: Ephemeral by Default

While Container is Active (~10 min after last request):

• ✅ Files in /workspace, /tmp, /home persist
• ✅ Background processes keep running
• ✅ Shell environment variables remain
• ✅ Session working directories preserved

When Container Goes Idle (after inactivity):

• ❌ ALL files deleted (entire filesystem reset)
• ❌ ALL processes terminated
• ❌ ALL shell state lost
• ⚠️ Next request creates fresh container from scratch

This is NOT like a traditional server. Sandboxes are ephemeral by design.

Handling Persistence

For Important Data: Use external storage

// Save to R2 before container goes idle
await sandbox.writeFile('/workspace/data.txt', content);
const fileData = await sandbox.readFile('/workspace/data.txt');
await env.R2.put('backup/data.txt', fileData);

// Restore on next request
const restored = await env.R2.get('backup/data.txt');
if (restored) {
  await sandbox.writeFile('/workspace/data.txt', await restored.text());
}

For Build Artifacts: Accept ephemerality or use caching

// Check if setup needed (handles cold starts)
const exists = await sandbox.readdir('/workspace/project').catch(() => null);
if (!exists) {
  await sandbox.gitCheckout(repoUrl, '/workspace/project');
  await sandbox.exec('npm install', { cwd: '/workspace/project' });
}
// Now safe to run build
await sandbox.exec('npm run build', { cwd: '/workspace/project' });

Session Management (Game-Changer for Chat Agents)

What Are Sessions?

Sessions are bash shell contexts within one sandbox. Think terminal tabs.

Key Properties:

• Each session has separate working directory
• Sessions share same filesystem
• Working directory persists across commands in same session
• Perfect for multi-step workflows

Pattern: Chat-Based Coding Agent

type ConversationState = {
  sandboxId: string;
  sessionId: string;
};

// First message: Create sandbox and session
const sandboxId = `user-${userId}`;
const sandbox = getSandbox(env.Sandbox, sandboxId);
const sessionId = await sandbox.createSession();

// Store in conversation state (database, KV, etc.)
await env.KV.put(`conversation:${conversationId}`, JSON.stringify({
  sandboxId,
  sessionId
}));

// Later messages: Reuse same session
const state = await env.KV.get(`conversation:${conversationId}`);
const { sandboxId, sessionId } = JSON.parse(state);
const sandbox = getSandbox(env.Sandbox, sandboxId);

// Commands run in same context
await sandbox.exec('cd /workspace/project', { session: sessionId });
await sandbox.exec('ls -la', { session: sessionId }); // Still in /workspace/project
await sandbox.exec('git status', { session: sessionId }); // Still in /workspace/project

Without Sessions (Common Mistake)

// ❌ WRONG: Each command runs in separate session
await sandbox.exec('cd /workspace/project');
await sandbox.exec('ls'); // NOT in /workspace/project (different session)

Pattern: Parallel Execution

const session1 = await sandbox.createSession();
const session2 = await sandbox.createSession();

// Run different tasks simultaneously
await Promise.all([
  sandbox.exec('python train_model.py', { session: session1 }),
  sandbox.exec('node generate_reports.js', { session: session2 })
]);

Sandbox Naming Strategies

Per-User Sandboxes (Persistent Workspace)

const sandbox = getSandbox(env.Sandbox, `user-${userId}`);

Pros: User's work persists while actively using (10 min idle time) Cons: Geographic lock-in (first request determines location) Use Cases: Interactive notebooks, IDEs, persistent workspaces

Per-Session Sandboxes (Fresh Each Time)

const sandboxId = `session-${Date.n