AI Agent Integration

This guide covers production patterns for deploying BoxLite as a sandboxed execution environment for AI agents. It assumes you’ve already worked through the Tutorials and know how to create boxes, run code, and transfer files.

Using BoxRun? If your agent communicates via HTTP, BoxRun’s REST API or Python SDK may be a simpler integration path. BoxRun handles sandbox lifecycle, file upload/download, and SSE streaming out of the box. See the AI agent patterns section in the BoxRun SDK docs.

Recommended Configuration

Workload-Type Reference

Workload	Image	CPUs	Memory	Disk	Notes
Code execution	`python:slim`	1	512 MiB	None	Ephemeral, fast startup
Data analysis	`python:slim`	2	2048 MiB	None	More memory for pandas/numpy
Web browsing	Use `BrowserBox`	2	2048 MiB	None	Chromium needs resources
Multi-tool agent	`python:slim`	2	1024 MiB	None	Balance cost vs. capability
Persistent env	`python:slim`	1	512 MiB	10 GB	State survives restarts

Starter Configuration

Python
Node.js

import boxlite
from boxlite.boxlite import SecurityOptions

options = boxlite.BoxOptions(
    image="python:slim",
    cpus=2,
    memory_mib=1024,
    working_dir="/workspace",
    security=SecurityOptions.maximum(),
)

import { JsBoxlite } from '@boxlite-ai/boxlite';

const runtime = JsBoxlite.withDefaultConfig();
const box = await runtime.create({
  image: 'python:slim',
  cpus: 2,
  memoryMib: 1024,
  workingDir: '/workspace',
});

Security Presets

SecurityOptions has three presets:

Preset	Jailer	Seccomp	Resource Limits	Use Case
`development()`	Off	Off	None	Debugging sandbox issues
`standard()`	On	On (Linux)	None	General workloads
`maximum()`	On	On (Linux)	`max_open_files=1024`, `max_file_size=1GiB`, `max_processes=100`	Untrusted AI code

For AI agents running untrusted code, use SecurityOptions.maximum():

Python
Node.js

from boxlite.boxlite import SecurityOptions

security = SecurityOptions.maximum()

# Customize if needed
security.max_open_files = 2048
security.network_enabled = False  # Disable network for strict isolation

The Node.js SDK enables jailer and seccomp by default. You configure resource limits directly in the box creation options:

import { JsBoxlite } from '@boxlite-ai/boxlite';

const runtime = JsBoxlite.withDefaultConfig();
const box = await runtime.create({
  image: 'python:slim',
  cpus: 2,
  memoryMib: 1024,
  maxOpenFiles: 2048,
  // Omit ports to prevent incoming connections
});

Concurrency Model

One Box, Multiple Executions (Recommended)

A single box can run many exec() calls. Each call spawns a new process inside the same VM. This avoids repeated VM boot overhead and is safe because the VM provides hardware isolation from the host.

Python
Node.js

import asyncio
import boxlite

async def main():
    runtime = boxlite.Boxlite.default()
    from boxlite.boxlite import SecurityOptions
    box = await runtime.create(boxlite.BoxOptions(
        image="python:slim",
        cpus=2,
        memory_mib=1024,
        security=SecurityOptions.maximum(),
    ))

    try:
        # Run agent tools concurrently in the same box
        results = await asyncio.gather(
            box.exec("python", ["-c", "print('task A')"]),
            box.exec("python", ["-c", "print('task B')"]),
            box.exec("python", ["-c", "print('task C')"]),
        )

        for execution in results:
            result = await execution.wait()
            print(f"Exit code: {result.exit_code}")
    finally:
        await box.stop()
        await runtime.remove(box.id)

import { JsBoxlite } from '@boxlite-ai/boxlite';

const runtime = JsBoxlite.withDefaultConfig();
const box = await runtime.create({
  image: 'python:slim',
  cpus: 2,
  memoryMib: 1024,
});

try {
  // Run agent tools concurrently in the same box
  const [execA, execB, execC] = await Promise.all([
    box.exec('python', ['-c', "print('task A')"]),
    box.exec('python', ['-c', "print('task B')"]),
    box.exec('python', ['-c', "print('task C')"]),
  ]);

  for (const execution of [execA, execB, execC]) {
    const result = await execution.wait();
    console.log(`Exit code: ${result.exitCode}`);
  }
} finally {
  await box.stop();
  await runtime.remove(box.id);
}

When to use: Most AI agent scenarios. Keeps VM boot cost to one-time.

One Box Per Agent

Use separate boxes when you need strict isolation between agents, different images, or independent resource limits.

Python
Node.js

async def run_isolated_agent(code: str, image: str = "python:slim"):
    """Each agent gets its own box."""
    async with boxlite.SimpleBox(image=image, memory_mib=512) as box:
        result = await box.exec("python", "-c", code)
        return result.stdout

async def main():
    agents = [
        run_isolated_agent("print('agent 1')"),
        run_isolated_agent("print('agent 2')", image="node:alpine"),
        run_isolated_agent("print('agent 3')"),
    ]
    results = await asyncio.gather(*agents)

import { JsBoxlite } from '@boxlite-ai/boxlite';

async function runIsolatedAgent(
  code: string,
  image: string = 'python:slim'
): Promise<string> {
  const runtime = JsBoxlite.withDefaultConfig();
  const box = await runtime.create({ image, memoryMib: 512 });

  try {
    const execution = await box.exec('python', ['-c', code]);
    const result = await execution.wait();
    return result.stdout;
  } finally {
    await box.stop();
    await runtime.remove(box.id);
  }
}

const results = await Promise.all([
  runIsolatedAgent("print('agent 1')"),
  runIsolatedAgent("print('agent 2')", 'node:alpine'),
  runIsolatedAgent("print('agent 3')"),
]);

When to use: Multi-tenant isolation, different language runtimes, or strict resource separation.

Timeout Handling and Zombie Prevention

The Problem

asyncio.wait_for() cancels the Python coroutine but does not kill the guest process. Without explicit cleanup, the process continues running inside the VM indefinitely.

The following pattern leaves a zombie process running inside the box:

# BAD: process keeps running inside the box after timeout
try:
    execution = await box.exec("python", ["-c", "import time; time.sleep(9999)"])
    result = await asyncio.wait_for(execution.wait(), timeout=5)
except asyncio.TimeoutError:
    print("Timed out")  # Process is still running in the VM!

Correct Pattern

Always kill the execution in the timeout handler:

Python
Node.js

async def exec_with_timeout(box, cmd, args=None, timeout=30):
    """Execute a command with proper timeout and cleanup."""
    execution = await box.exec(cmd, args or [])
    try:
        result = await asyncio.wait_for(execution.wait(), timeout=timeout)
        return result
    except asyncio.TimeoutError:
        await execution.kill()
        raise

async function execWithTimeout(box, cmd, args = [], timeoutMs = 30000) {
  const execution = await box.exec(cmd, args);
  let timer;
  try {
    const result = await Promise.race([
      execution.wait(),
      new Promise((_, reject) => {
        timer = setTimeout(
          () => reject(new Error('Execution timed out')),
          timeoutMs
        );
      }),
    ]);
    clearTimeout(timer);
    return result;
  } catch (error) {
    clearTimeout(timer);
    await execution.kill();
    throw error;
  }
}

Defensive Helper

For maximum safety, combine timeout handling with a try/finally block:

Python
Node.js

async def safe_exec(box, cmd, args=None, timeout=30):
    """Execute with timeout, guaranteed process cleanup."""
    execution = await box.exec(cmd, args or [])
    try:
        result = await asyncio.wait_for(execution.wait(), timeout=timeout)
        return result
    except asyncio.TimeoutError:
        try:
            await execution.kill()
        except Exception:
            pass  # Best-effort kill
        raise
    except Exception:
        try:
            await execution.kill()
        except Exception:
            pass  # Best-effort kill on any failure
        raise

async function safeExec(box, cmd, args = [], timeoutMs = 30000) {
  const execution = await box.exec(cmd, args);
  let timer;
  try {
    const result = await Promise.race([
      execution.wait(),
      new Promise((_, reject) => {
        timer = setTimeout(
          () => reject(new Error('Execution timed out')),
          timeoutMs
        );
      }),
    ]);
    clearTimeout(timer);
    return result;
  } catch (error) {
    clearTimeout(timer);
    try {
      await execution.kill();
    } catch {
      // Best-effort kill
    }
    throw error;
  }
}

Security Boundaries

SecurityOptions Fields

Field	Type	Description
`jailer_enabled`	`bool`	OS-level sandbox (seccomp on Linux, sandbox-exec on macOS)
`seccomp_enabled`	`bool`	Syscall filtering (Linux only)
`max_open_files`	`int \| None`	Limit open file descriptors
`max_file_size`	`int \| None`	Maximum file size in bytes
`max_processes`	`int \| None`	Maximum number of processes
`max_memory`	`int \| None`	Maximum virtual memory in bytes
`max_cpu_time`	`int \| None`	Maximum CPU time in seconds
`network_enabled`	`bool`	Allow network access from sandbox (macOS only)
`close_fds`	`bool`	Close inherited file descriptors

Network Isolation

To prevent an agent from accessing the network:

Python
Node.js

from boxlite.boxlite import SecurityOptions

security = SecurityOptions.maximum()
security.network_enabled = False

options = boxlite.BoxOptions(
    image="python:slim",
    security=security,
    # No ports= means no incoming connections either
)

import { JsBoxlite } from '@boxlite-ai/boxlite';

const runtime = JsBoxlite.withDefaultConfig();
const box = await runtime.create({
  image: 'python:slim',
  // Do not pass ports — no incoming connections
  // On Linux, the VM runs in an isolated network namespace by default
});

// Verify: this should fail if network is isolated
const execution = await box.exec('ping', ['-c', '1', '8.8.8.8']);
const result = await execution.wait();
console.log(`Network test exit code: ${result.exitCode}`);

The Node.js SDK does not expose a networkEnabled toggle. On Linux, network isolation is enforced by the runtime’s network namespace. On macOS, use the Python SDK’s SecurityOptions.network_enabled for explicit control.

In the Python bindings, network_enabled is currently a macOS-only control. On Linux and other platforms, network isolation is typically enforced by the container/runtime networking configuration (for example, running in an isolated network namespace and not publishing ports), and network_enabled may not itself hard-disable all outbound connectivity.

Resource Limits as Security Boundaries

Resource limits prevent a rogue agent from consuming all host resources:

options = boxlite.BoxOptions(
    image="python:slim",
    cpus=1,             # Cap CPU usage
    memory_mib=512,     # Hard memory limit
    security=SecurityOptions.maximum(),
)

Memory Limits and OOM

The memory_mib setting is a hard limit enforced by the hypervisor. When a guest process exceeds this limit, the Linux OOM killer terminates the offending process inside the VM — but the box itself stays running. This means you can detect OOM and retry or report the failure. How to detect OOM:

The process exit code will be 137 (128 + SIGKILL)
stderr may contain Killed or Out of memory

result = await box.exec("python", ["-c", "x = bytearray(2**30)"])
completed = await result.wait()

if completed.exit_code == 137:
    print("Process was OOM-killed")
    print(f"stderr: {completed.stderr}")

OOM kills the guest process, not the box. The box remains running and can accept new exec() calls. If your agent needs to detect and handle OOM, check for exit code 137 after each execution.

Terminal Resizing

When running interactive TTY sessions (e.g., an AI agent controlling a shell), use resize_tty() to set the terminal dimensions. This ensures proper line wrapping and avoids garbled output from programs that query terminal size.

Python
Node.js

runtime = boxlite.Boxlite.default()
box = await runtime.create(boxlite.BoxOptions(image="alpine:latest"))

# Start a shell with TTY
execution = await box.exec("sh", tty=True)

# Set terminal size to 40 rows x 120 columns
await execution.resize_tty(40, 120)

# Send commands via stdin
stdin = execution.stdin()
await stdin.send_input(b"ls -la\n")

# Read output
stdout = execution.stdout()
async for line in stdout:
    print(line)

import { JsBoxlite } from '@boxlite-ai/boxlite';

const runtime = JsBoxlite.withDefaultConfig();
const box = await runtime.create({ image: 'alpine:latest' });

// Start a shell with TTY
const execution = await box.exec('sh', [], { tty: true });

// Set terminal size to 40 rows x 120 columns
await execution.resizeTty(40, 120);

// Send commands via stdin
const stdin = execution.stdin();
await stdin.sendInput(Buffer.from('ls -la\n'));

// Read output
const stdout = execution.stdout();
for await (const line of stdout) {
  console.log(line);
}

resize_tty() / resizeTty() only works on executions started with tty=True / { tty: true }. Calling it on a non-TTY execution returns an error.

Complete Example

Putting it all together: security configuration, concurrent execution with timeouts, and cleanup.

Python
Node.js

import asyncio
import boxlite
from boxlite.boxlite import SecurityOptions


async def safe_exec(box, cmd, args=None, timeout=30):
    """Execute with timeout and guaranteed process cleanup."""
    execution = await box.exec(cmd, args or [])
    try:
        result = await asyncio.wait_for(execution.wait(), timeout=timeout)
        return result
    except asyncio.TimeoutError:
        try:
            await execution.kill()
        except Exception:
            pass
        raise


async def main():
    runtime = boxlite.Boxlite.default()

    # Configure box with security and resource limits
    box = await runtime.create(boxlite.BoxOptions(
        image="python:slim",
        cpus=2,
        memory_mib=1024,
        working_dir="/workspace",
        security=SecurityOptions.maximum(),
    ))

    try:
        # Run with timeout protection
        result = await safe_exec(
            box,
            "python",
            ["-c", "print('hello from secure sandbox')"],
            timeout=60,
        )
        print(f"Exit code: {result.exit_code}")

        # Run concurrent tasks safely
        tasks = [
            safe_exec(box, "python", ["-c", "print('task 1')"], timeout=10),
            safe_exec(box, "python", ["-c", "print('task 2')"], timeout=10),
        ]
        results = await asyncio.gather(*tasks, return_exceptions=True)

        for i, r in enumerate(results):
            if isinstance(r, Exception):
                print(f"Task {i} failed: {r}")
            else:
                print(f"Task {i} exit code: {r.exit_code}")

    finally:
        await box.stop()
        await runtime.remove(box.id)


asyncio.run(main())

import { JsBoxlite } from '@boxlite-ai/boxlite';

async function safeExec(box, cmd, args = [], timeoutMs = 30000) {
  const execution = await box.exec(cmd, args);
  let timer;
  try {
    const result = await Promise.race([
      execution.wait(),
      new Promise((_, reject) => {
        timer = setTimeout(
          () => reject(new Error('Execution timed out')),
          timeoutMs
        );
      }),
    ]);
    clearTimeout(timer);
    return result;
  } catch (error) {
    clearTimeout(timer);
    try { await execution.kill(); } catch { /* best-effort */ }
    throw error;
  }
}

const runtime = JsBoxlite.withDefaultConfig();
const box = await runtime.create({
  image: 'python:slim',
  cpus: 2,
  memoryMib: 1024,
  workingDir: '/workspace',
});

try {
  // Run with timeout protection
  const result = await safeExec(
    box,
    'python',
    ['-c', "print('hello from secure sandbox')"],
    60000
  );
  console.log(`Exit code: ${result.exitCode}`);

  // Run concurrent tasks safely
  const results = await Promise.allSettled([
    safeExec(box, 'python', ['-c', "print('task 1')"], 10000),
    safeExec(box, 'python', ['-c', "print('task 2')"], 10000),
  ]);

  results.forEach((r, i) => {
    if (r.status === 'rejected') {
      console.log(`Task ${i} failed: ${r.reason}`);
    } else {
      console.log(`Task ${i} exit code: ${r.value.exitCode}`);
    }
  });
} finally {
  await box.stop();
  await runtime.remove(box.id);
}

Getting Started

Tutorials

Guides

SDK Reference

Architecture

Resources

Development

Recommended Configuration

Workload-Type Reference

Starter Configuration

Security Presets

Concurrency Model

One Box, Multiple Executions (Recommended)

One Box Per Agent

Timeout Handling and Zombie Prevention

The Problem

Correct Pattern

Defensive Helper

Security Boundaries

SecurityOptions Fields

Network Isolation

Resource Limits as Security Boundaries

Memory Limits and OOM

Terminal Resizing

Complete Example

See also

Getting Started

Tutorials

Guides

SDK Reference

Architecture

Resources

Development

​Recommended Configuration

​Workload-Type Reference

​Starter Configuration

​Security Presets

​Concurrency Model

​One Box, Multiple Executions (Recommended)

​One Box Per Agent

​Timeout Handling and Zombie Prevention

​The Problem

​Correct Pattern

​Defensive Helper

​Security Boundaries

​SecurityOptions Fields

​Network Isolation

​Resource Limits as Security Boundaries

​Memory Limits and OOM

​Terminal Resizing

​Complete Example

​See also

Recommended Configuration

Workload-Type Reference

Starter Configuration

Security Presets

Concurrency Model

One Box, Multiple Executions (Recommended)

One Box Per Agent

Timeout Handling and Zombie Prevention

The Problem

Correct Pattern

Defensive Helper

Security Boundaries

SecurityOptions Fields

Network Isolation

Resource Limits as Security Boundaries

Memory Limits and OOM

Terminal Resizing

Complete Example

See also