Connect an LLM

Vibengine sandboxes are designed to be the execution environment for AI agents. Here's how to connect your LLM to a sandbox for code execution.

Basic Pattern

The typical workflow is:

Your LLM generates code or commands
Your application sends them to a Vibengine sandbox
The sandbox executes and returns results
Results are fed back to the LLM

import { Sandbox } from 'vibengine'

// Create a long-running sandbox for the agent session
const sandbox = await Sandbox.create({ timeoutMs: 600_000 }) // 10 min

async function executeCode(code) {
  // Write the code to a file
  await sandbox.files.write('/home/user/script.py', code)

  // Execute it
  const result = await sandbox.commands.run('python3 /home/user/script.py')

  return {
    stdout: result.stdout,
    stderr: result.stderr,
    exitCode: result.exitCode,
  }
}

// Example: Execute LLM-generated code
const code = `
import math
print(f"Pi is approximately {math.pi:.10f}")
`
const result = await executeCode(code)
console.log(result.stdout) // "Pi is approximately 3.1415926536"

await sandbox.kill()

from vibengine import Sandbox

# Create a long-running sandbox for the agent session
sandbox = Sandbox(timeout=600)  # 10 min

def execute_code(code: str) -> dict:
    # Write the code to a file
    sandbox.files.write('/home/user/script.py', code)

    # Execute it
    result = sandbox.commands.run('python3 /home/user/script.py')

    return {
        'stdout': result.stdout,
        'stderr': result.stderr,
        'exit_code': result.exit_code,
    }

# Example: Execute LLM-generated code
code = """
import math
print(f"Pi is approximately {math.pi:.10f}")
"""
result = execute_code(code)
print(result['stdout'])  # "Pi is approximately 3.1415926536"

sandbox.kill()

Using with OpenAI

import { Sandbox } from 'vibengine'
import OpenAI from 'openai'

const openai = new OpenAI()
const sandbox = await Sandbox.create({ timeoutMs: 600_000 })

const tools = [{
  type: 'function',
  function: {
    name: 'execute_code',
    description: 'Execute Python code in a sandbox',
    parameters: {
      type: 'object',
      properties: {
        code: { type: 'string', description: 'Python code to execute' }
      },
      required: ['code']
    }
  }
}]

// Chat with tool calling
const response = await openai.chat.completions.create({
  model: 'gpt-4o',
  messages: [{ role: 'user', content: 'Calculate the first 10 fibonacci numbers' }],
  tools,
})

// Handle tool calls
for (const toolCall of response.choices[0].message.tool_calls || []) {
  if (toolCall.function.name === 'execute_code') {
    const { code } = JSON.parse(toolCall.function.arguments)
    await sandbox.files.write('/home/user/script.py', code)
    const result = await sandbox.commands.run('python3 /home/user/script.py')
    console.log('Output:', result.stdout)
  }
}

await sandbox.kill()

from vibengine import Sandbox
from openai import OpenAI
import json

client = OpenAI()
sandbox = Sandbox(timeout=600)

tools = [{
    "type": "function",
    "function": {
        "name": "execute_code",
        "description": "Execute Python code in a sandbox",
        "parameters": {
            "type": "object",
            "properties": {
                "code": {"type": "string", "description": "Python code to execute"}
            },
            "required": ["code"]
        }
    }
}]

response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Calculate the first 10 fibonacci numbers"}],
    tools=tools,
)

for tool_call in response.choices[0].message.tool_calls or []:
    if tool_call.function.name == "execute_code":
        args = json.loads(tool_call.function.arguments)
        sandbox.files.write('/home/user/script.py', args['code'])
        result = sandbox.commands.run('python3 /home/user/script.py')
        print('Output:', result.stdout)

sandbox.kill()

The same pattern works with any LLM provider — Claude, Gemini, Llama, etc. The key is to use Vibengine as the execution backend for tool calls.

Connect an LLM

Connect an LLM

Basic Pattern

Using with OpenAI

On this page