DEV Community

Ozor
Ozor

Posted on

How to Execute Untrusted Code Safely with a Sandbox API (Python, JS, Bash)

#ai #security #python #javascript

If you're building an AI agent, coding assistant, or educational platform, you'll eventually need to execute user-submitted code. And if you've ever tried to run eval() on untrusted input, you know that's a one-way ticket to getting pwned.

The safe way? Run code in an isolated sandbox — no filesystem access, no network, strict timeouts, resource limits. But setting up Docker containers, gVisor, or Firecracker VMs is a pain.

In this post, I'll show you how to execute Python, JavaScript, TypeScript, and Bash code safely using a free sandbox API — zero infrastructure, one HTTP call.

Why You Need a Code Sandbox

Common use cases:

  • AI agents that write and test code (LLM tool use)
  • Online code editors and playgrounds
  • Coding challenges and interview platforms
  • Education platforms where students run exercises
  • CI/CD pipelines that need quick code validation

The risk of running code directly:

# Never do this
exec(user_input)  # RCE in 3 characters
Enter fullscreen mode Exit fullscreen mode 
// Also never do this
eval(userCode); // Full access to your Node.js process
Enter fullscreen mode Exit fullscreen mode

The API: One Endpoint, Four Languages

Here's the simplest possible code execution:

curl -X POST https://agent-gateway-kappa.vercel.app/v1/agent-coderunner/execute \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "language": "python",
    "code": "print(sum(range(1, 101)))"
  }'
Enter fullscreen mode Exit fullscreen mode

Response:

{
  "id": "eb44ad44-b024-4a90-a1dd-1b6adc3f9f93",
  "language": "python",
  "exitCode": 0,
  "stdout": "5050",
  "stderr": "",
  "duration": 13,
  "timedOut": false
}
Enter fullscreen mode Exit fullscreen mode

Get a free API key (200 credits, no signup):

curl -X POST https://agent-gateway-kappa.vercel.app/api/keys/create
Enter fullscreen mode Exit fullscreen mode

Running Python

Python 3.12, standard library included:

const response = await fetch('https://agent-gateway-kappa.vercel.app/v1/agent-coderunner/execute', {
  method: 'POST',
  headers: {
    'Content-Type': 'application/json',
    'Authorization': 'Bearer YOUR_API_KEY'
  },
  body: JSON.stringify({
    language: 'python',
    code: `
import json
import math

data = [math.sqrt(i) for i in range(1, 11)]
print(json.dumps({"roots": [round(x, 3) for x in data]}))
`
  })
});

const result = await response.json();
console.log(result.stdout);
// {"roots": [1.0, 1.414, 1.732, 2.0, 2.236, 2.449, 2.646, 2.828, 3.0, 3.162]}
Enter fullscreen mode Exit fullscreen mode

Running JavaScript (Node.js 20) 

import requests

result = requests.post(
    'https://agent-gateway-kappa.vercel.app/v1/agent-coderunner/execute',
    headers={'Authorization': 'Bearer YOUR_API_KEY'},
    json={
        'language': 'javascript',
        'code': '''
const crypto = require("crypto");
const hash = crypto.createHash("sha256").update("hello world").digest("hex");
console.log(JSON.stringify({ input: "hello world", sha256: hash }));
'''
    }
).json()

print(f"Exit code: {result['exitCode']}")
print(f"Output: {result['stdout']}")
Enter fullscreen mode Exit fullscreen mode

Running Bash

For quick scripting and system-level checks:

curl -X POST https://agent-gateway-kappa.vercel.app/v1/agent-coderunner/execute \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "language": "bash",
    "code": "echo \"CPU cores: $(nproc)\"\necho \"Memory: $(free -h | awk '/Mem:/ {print $2}')\"\necho \"Uptime: $(uptime -p)\""
  }'
Enter fullscreen mode Exit fullscreen mode

Building an AI Agent Tool

The killer use case: give your LLM the ability to write code and test it. Here's an OpenAI function-calling integration:

const tools = [{
  type: 'function',
  function: {
    name: 'execute_code',
    description: 'Execute Python or JavaScript code in a secure sandbox',
    parameters: {
      type: 'object',
      properties: {
        language: { type: 'string', enum: ['python', 'javascript'] },
        code: { type: 'string', description: 'Code to execute' }
      },
      required: ['language', 'code']
    }
  }
}];

async function executeCode(language, code) {
  const res = await fetch('https://agent-gateway-kappa.vercel.app/v1/agent-coderunner/execute', {
    method: 'POST',
    headers: {
      'Content-Type': 'application/json',
      'Authorization': `Bearer ${process.env.GATEWAY_API_KEY}`
    },
    body: JSON.stringify({ language, code })
  });

  const result = await res.json();

  if (result.exitCode !== 0) {
    return `Error (exit ${result.exitCode}):\n${result.stderr}`;
  }
  return result.stdout;
}
Enter fullscreen mode Exit fullscreen mode

Now your AI agent can:

  1. Write code to solve a problem
  2. Execute it in a sandbox
  3. Read the output
  4. Fix errors and retry

No Docker, no VMs, no infrastructure.

Handling Errors Gracefully

The API returns structured error info you can act on:

async function safeExecute(language, code) {
  const result = await executeCode(language, code);

  if (result.timedOut) {
    return { error: 'Code execution timed out (10s limit)' };
  }

  if (result.exitCode !== 0) {
    return {
      error: 'Runtime error',
      stderr: result.stderr,
      exitCode: result.exitCode
    };
  }

  return {
    output: result.stdout,
    duration: `${result.duration}ms`
  };
}
Enter fullscreen mode Exit fullscreen mode

What the Sandbox Blocks

Each execution runs in an isolated environment:

Feature Status
Filesystem read/write Restricted
Network access Blocked
Process spawning Limited
Execution time 10s timeout
Memory Capped
Stdout Truncated at 64KB

This means malicious code like import os; os.system("rm -rf /") or while True: pass won't harm anything — it'll either be blocked or time out.

Comparing Approaches

Approach Setup Time Cost Isolation Latency
eval() / exec() 0 min Free None (dangerous) <1ms
Docker containers Hours Server costs Good 500ms+ cold start
AWS Lambda 30 min Pay per invoke Great 100-500ms
Firecracker/gVisor Days Server costs Excellent 50-200ms
Sandbox API 1 min Free (200 credits) Good ~15ms

Try It Now

Get your API key and run your first sandboxed code:

# Get a free key
API_KEY=$(curl -s -X POST https://agent-gateway-kappa.vercel.app/api/keys/create | jq -r '.key')

# Run Python
curl -s -X POST https://agent-gateway-kappa.vercel.app/v1/agent-coderunner/execute \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $API_KEY" \
  -d '{"language":"python","code":"print(\"It works!\")"}' | jq .

# Run JavaScript
curl -s -X POST https://agent-gateway-kappa.vercel.app/v1/agent-coderunner/execute \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $API_KEY" \
  -d '{"language":"javascript","code":"console.log(Array.from({length:10}, (_,i) => i*i))"}' | jq .
Enter fullscreen mode Exit fullscreen mode

200 free credits, no signup required. The API also gives you 40+ other services (crypto prices, IP geolocation, screenshots, DNS lookup) through the same key.

Full API docs | Interactive playground | Get API key

Top comments (0)