sandbox-ai

Lightweight AI code execution sandbox. Run untrusted code safely with resource limits, timeout protection, and output capture.

npx sandbox-ai "print('hello world')" --lang python

Why

AI agents generate code. Running that code directly is dangerous. sandbox-ai provides isolated execution with:

• Resource limits (CPU, memory, time)
• No network access by default
• No filesystem access outside sandbox
• Automatic cleanup
• Structured output (stdout, stderr, exit code, timing)

Quick Start

# Run Python
npx sandbox-ai "import math; print(math.pi)" --lang python

# Run JavaScript
npx sandbox-ai "console.log(Array.from({length:10}, (_,i) => i*i))" --lang node

# Run with timeout
npx sandbox-ai "while True: pass" --lang python --timeout 5

# Run from file
npx sandbox-ai --file solution.py --lang python

# Get JSON output
npx sandbox-ai "print(42)" --lang python --json

Output

$ npx sandbox-ai "for i in range(5): print(f'fib({i})')" --lang python

  sandbox-ai v1.0.0

  Language: python
  Timeout:  30s

  ─── Output ───
  fib(0)
  fib(1)
  fib(2)
  fib(3)
  fib(4)
  ──────────────

  ✓ Exit code: 0
  ⏱ Time: 0.12s

JSON Output

{
  "stdout": "fib(0)\nfib(1)\nfib(2)\nfib(3)\nfib(4)\n",
  "stderr": "",
  "exitCode": 0,
  "elapsed_ms": 120,
  "language": "python",
  "timedOut": false
}

Supported Languages

Language	Flag	Runtime
Python	--lang python	python3
JavaScript	--lang node	node
TypeScript	--lang ts	tsx / ts-node
Bash	--lang bash	bash
Go	--lang go	go run
Ruby	--lang ruby	ruby

Safety Features

Feature	Default	Description
Timeout	30s	Kill process after N seconds
No network	On	Block outbound connections
Temp dir	On	Run in isolated temp directory
Cleanup	On	Remove temp files after execution
Max output	100KB	Truncate output at limit

API Usage

import { sandbox } from 'sandbox-ai';

const result = await sandbox({
  code: 'print("hello")',
  language: 'python',
  timeout: 10000,
  allowNetwork: false,
});

console.log(result.stdout);  // "hello\n"
console.log(result.exitCode); // 0
console.log(result.elapsed_ms); // 45

Use Cases

• AI agent code execution — Run LLM-generated code safely
• Code playground — Build web-based code runners
• Testing — Execute test cases in isolation
• Education — Safe student code execution
• CI/CD — Run arbitrary scripts with limits

License

MIT