Deno Sandbox

Deno Sandbox

by Deno
MIT
Repository Docs
markdown Development
denosandboxsecuritycode-execution

Summary

Run untrusted or model-generated code in an isolated environment using the @deno/sandbox SDK.

Features

  • Execute untrusted or AI-generated code safely
  • Covers the @deno/sandbox SDK
  • Permission-scoped isolation for agent tools

Install This Skill

Add this skill to your favorite AI agent in a few steps.

Any AI agent

This skill is plain instructions — it works with any assistant that accepts custom instructions or system prompts.

  1. Copy the skill content with the button below.
  2. Paste it into your agent's instruction file or system prompt (for example AGENTS.md, .cursorrules, or a custom instructions field).
  3. Ask the agent to apply the skill whenever the task matches.

Skill Content

Markdown Content

Copy this content and use it with your preferred AI agent

# Deno Sandboxes

## Overview

Deno Sandboxes provide secure, isolated environments for running untrusted code.
Each sandbox runs in its own Linux microVM (using Firecracker, the same
technology as AWS Lambda) with a separate filesystem, network, and process
space. This makes them ideal for code playgrounds, AI agent tool execution, and
multi-tenant applications.

Reference: https://deno.com/deploy/sandboxes

## When to Use Sandboxes

Use Deno Sandboxes when you need to:

- Run user-submitted code safely
- Execute AI-generated code
- Build code playground platforms
- Create multi-tenant code execution environments
- Run automated tests in isolation

## Scope Boundaries

This skill applies **only** to Deno Sandbox (`@deno/sandbox`) questions. Follow
these rules:

- If the user asks about **Docker containers, Node.js vm2/vm, Web Workers, AWS
  Lambda isolation, or any other isolation technology**, answer using that
  technology directly. Do not suggest Deno Sandboxes as an alternative.
- Do not include `@deno/sandbox` imports, `Sandbox.create()`, or
  `sandbox.spawn()` in responses about other isolation technologies.
- Only recommend Deno Sandboxes when the user explicitly asks about them or is
  building code execution features in a Deno project.

## Getting Started

### Installation

```bash
deno add jsr:@deno/sandbox
```

### Basic Usage

```typescript
import { Sandbox } from "@deno/sandbox";

// Create a sandbox (auto-disposed when scope ends)
await using sandbox = await Sandbox.create();

// Run a command
const child = await sandbox.spawn("echo", { args: ["Hello from sandbox!"] });
const output = await child.output();

console.log(new TextDecoder().decode(output.stdout));
// Output: Hello from sandbox!
```

## Core Concepts

### Sandbox Lifecycle

Sandboxes are resources that must be disposed when done. **Always** use
`await using` for automatic cleanup:

```typescript
await using sandbox = await Sandbox.create();
// Sandbox is automatically destroyed when this scope ends
```

CRITICAL: Never show `const sandbox = await Sandbox.create()` without
`await using`. Always use the `await using` pattern for sandbox creation. Do not
show manual disposal alternatives.

### Running Processes

The `spawn` method runs commands inside the sandbox:

```typescript
const child = await sandbox.spawn("deno", {
  args: ["run", "script.ts"],
  stdin: "piped", // Enable stdin
  stdout: "piped", // Capture stdout
  stderr: "piped", // Capture stderr
});

// Wait for completion and get output
const output = await child.output();
console.log("Exit code:", output.code);
console.log("Stdout:", new TextDecoder().decode(output.stdout));
console.log("Stderr:", new TextDecoder().decode(output.stderr));
```

### Streaming I/O

For interactive processes or long-running commands:

```typescript
const child = await sandbox.spawn("deno", {
  args: ["repl"],
  stdin: "piped",
  stdout: "piped",
});

// Write to stdin
const writer = child.stdin!.getWriter();
await writer.write(new TextEncoder().encode("console.log('Hello')\n"));
await writer.close();

// Read from stdout
const reader = child.stdout!.getReader();
while (true) {
  const { done, value } = await reader.read();
  if (done) break;
  console.log(new TextDecoder().decode(value));
}
```

### Killing Processes

```typescript
const child = await sandbox.spawn("sleep", { args: ["60"] });

// Kill with SIGTERM (default)
await child.kill();

// Or with specific signal
await child.kill("SIGKILL");

// Wait for exit
const status = await child.status;
console.log("Exited with signal:", status.signal);
```

## Common Patterns

### Running User Code Safely

```typescript
import { Sandbox } from "@deno/sandbox";

async function runUserCode(code: string): Promise<string> {
  await using sandbox = await Sandbox.create();

  // Write user code to a file in the sandbox
  await sandbox.fs.writeFile("/tmp/user_code.ts", code);

  // Run with restricted permissions
  const child = await sandbox.spawn("deno", {
    args: [
      "run",
      "--allow-none", // No permissions
      "/tmp/user_code.ts",
    ],
    stdout: "piped",
    stderr: "piped",
  });

  const output = await child.output();

  if (output.code !== 0) {
    throw new Error(new TextDecoder().decode(output.stderr));
  }

  return new TextDecoder().decode(output.stdout);
}
```

### Code Playground

```typescript
import { Sandbox } from "@deno/sandbox";

interface ExecutionResult {
  success: boolean;
  output: string;
  error?: string;
  executionTime: number;
}

async function executePlayground(code: string): Promise<ExecutionResult> {
  const start = performance.now();

  await using sandbox = await Sandbox.create();

  await sandbox.fs.writeFile("/playground/main.ts", code);

  const child = await sandbox.spawn("deno", {
    args: ["run", "--allow-net", "/playground/main.ts"],
    stdout: "piped",
    stderr: "piped",
  });

  const output = await child.output();
  const executionTime = performance.now() - start;

  return {
    success: output.code === 0,
    output: new TextDecoder().decode(output.stdout),
    error: output.code !== 0
      ? new TextDecoder().decode(output.stderr)
      : undefined,
    executionTime,
  };
}
```

### AI Agent Tool Execution

```typescript
import { Sandbox } from "@deno/sandbox";

async function executeAgentTool(
  toolCode: string,
  input: unknown,
): Promise<unknown> {
  await using sandbox = await Sandbox.create();

  // Create a wrapper that handles input/output
  const wrapper = `
    const input = ${JSON.stringify(input)};
    const tool = await import("/tool.ts");
    const result = await tool.default(input);
    console.log(JSON.stringify(result));
  `;

  await sandbox.fs.writeFile("/tool.ts", toolCode);
  await sandbox.fs.writeFile("/run.ts", wrapper);

  const child = await sandbox.spawn("deno", {
    args: ["run", "--allow-net", "/run.ts"],
    stdout: "piped",
    stderr: "piped",
  });

  const output = await child.output();

  if (output.code !== 0) {
    throw new Error(new TextDecoder().decode(output.stderr));
  }

  return JSON.parse(new TextDecoder().decode(output.stdout));
}
```

## Sandbox Features

### Resource Configuration

Sandboxes have configurable resources:

- **Default:** 2 vCPUs, 512MB memory, 10GB disk
- Startup time: Under 200ms

### What's Included

Each sandbox comes with:

- TypeScript/JavaScript runtime (Deno)
- Full Linux environment
- Network access (can be restricted)
- Temporary filesystem

### Security Features

- **Firecracker microVMs** - Same technology as AWS Lambda
- **Full isolation** - Separate kernel, filesystem, network
- **No data leakage** - Sandboxes can't access host system
- **Enforced policies** - Control outbound connections

## Deploying Sandboxes

Sandboxes can be deployed directly to Deno Deploy:

```bash
deno deploy --prod
```

The sandbox SDK works seamlessly in the Deno Deploy environment.

## API Reference

For the complete API, run:

```bash
deno doc jsr:@deno/sandbox
```

Key classes:

- `Sandbox` - Main class for creating/managing sandboxes
- `ChildProcess` - Represents a running process
- `Client` - For managing Deploy resources (apps, volumes)

## Quick Reference

| Task           | Code                                           |
| -------------- | ---------------------------------------------- |
| Create sandbox | `await using sandbox = await Sandbox.create()` |
| Run command    | `sandbox.spawn("cmd", { args: [...] })`        |
| Get output     | `const output = await child.output()`          |
| Write file     | `await sandbox.fs.writeFile(path, content)`    |
| Read file      | `await sandbox.fs.readFile(path)`              |
| Kill process   | `await child.kill()`                           |
| Check status   | `const status = await child.status`            |

## Common Mistakes

**Forgetting automatic disposal**

```typescript
// ❌ Wrong - always use "await using" for sandbox creation
// Never write: const sandbox = await Sandbox.create() without "await using"

// ✅ Correct - use "await using" for automatic cleanup
await using sandbox = await Sandbox.create();
await sandbox.spawn("echo", { args: ["hello"] });
// sandbox automatically disposed when scope ends
```

**Giving user code too many permissions**

```typescript
// ❌ Wrong - gives untrusted code full access
const child = await sandbox.spawn("deno", {
  args: ["run", "--allow-all", "/tmp/user_code.ts"],
});

// ✅ Correct - restrict permissions to what's needed
const child = await sandbox.spawn("deno", {
  args: ["run", "--allow-none", "/tmp/user_code.ts"], // No permissions
});

// Or if network is truly needed:
const child = await sandbox.spawn("deno", {
  args: ["run", "--allow-net", "/tmp/user_code.ts"], // Only network
});
```

**Not handling process output properly**

```typescript
// ❌ Wrong - forgetting to pipe stdout/stderr
const child = await sandbox.spawn("deno", { args: ["run", "script.ts"] });
const output = await child.output();
// output.stdout is empty because we didn't pipe it!

// ✅ Correct - pipe the streams you need
const child = await sandbox.spawn("deno", {
  args: ["run", "script.ts"],
  stdout: "piped",
  stderr: "piped",
});
const output = await child.output();
console.log(new TextDecoder().decode(output.stdout));
```

**Not setting timeouts for user code execution**

```typescript
// ❌ Wrong - user code could run forever
const child = await sandbox.spawn("deno", {
  args: ["run", "/tmp/user_code.ts"],
});
await child.output(); // Could hang indefinitely

// ✅ Correct - implement timeout handling
const child = await sandbox.spawn("deno", {
  args: ["run", "/tmp/user_code.ts"],
  stdout: "piped",
  stderr: "piped",
});

// Set a timeout to kill the process
const timeoutId = setTimeout(() => child.kill(), 5000); // 5 second limit

try {
  const output = await child.output();
  return output;
} finally {
  clearTimeout(timeoutId);
}
```

**Trusting sandbox output without validation**

```typescript
// ❌ Wrong - directly using untrusted output as code
const result = await runUserCode(code);
// Never execute or inject untrusted output!

// ✅ Correct - validate and sanitize output
const result = await runUserCode(code);
try {
  const parsed = JSON.parse(result); // Parse as data, not code
  if (isValidResponse(parsed)) {
    return parsed;
  }
} catch {
  throw new Error("Invalid response from sandbox");
}
```

Description

Deno's permission model makes it a natural fit for running code you do not trust, and the @deno/sandbox SDK formalises that into an API.

This official skill covers building features that execute untrusted input: creating sandboxes with scoped permissions, passing code and data in, capturing output and errors, enforcing resource limits, and tearing down cleanly.

The obvious application is AI agents that write and run code, where the sandbox is the difference between a useful tool and an arbitrary-code-execution vulnerability. It also applies to user-submitted scripts, plugin systems, and educational platforms. Published by the Deno team.

Related Skills

Browse all skills →