Control internet access for AI agent sandboxes

You can decide whether a new Celesto computer can reach the internet. Turn internet access off when you want the computer to work only from files and tools already on disk, for example when an agent runs untrusted code, processes sensitive input that must not leave the sandbox, or reproduces a build with no outside package fetches. The computer can still run commands and provide a terminal for your agent. Set the choice when you create a computer. It stays with that computer across stop, start, and restore, so create a new computer when a task needs a different setting.

Network Control requires Celesto SDK 0.0.13 or later for Python and 0.1.5 or later for TypeScript. The --no-internet CLI flag ships in the same Python release.

Choose an internet setting

Setting	What your computer can do
`open`	Connect to the internet. This is the default.
`off`	Run commands and use terminal sessions with outbound internet access turned off.

Network Control currently supports these two settings. Domain and IP allowlists are not available.

Create an offline computer

Pass network_policy={"mode": "off"} when you create a computer with Python, or networkPolicy: { mode: "off" } with TypeScript. The CLI provides the same choice through --no-internet.

from celesto import Computer


computer = Computer(network_policy={"mode": "off"})
try:
    result = computer.run("uname -a")
    print(result["stdout"].strip())
finally:
    computer.delete()

import { Computer } from "@celestoai/sdk";


const computer = await Computer.create({ networkPolicy: { mode: "off" } });

try {
  const result = await computer.run("uname -a");
  console.log(result.stdout.trim());
} finally {
  await computer.delete();
}

celesto computer create --no-internet

The Python and TypeScript examples print information about the computer’s operating system. Commands and terminal sessions remain available while outbound internet access is off.

Keep the setting with the computer

Internet access is fixed when the computer is created. Stopping, starting, and restoring a computer keeps the same setting. This makes an offline workflow predictable when an agent resumes work later. You can read the selected setting from the computer response:

from celesto import Computer


computer = Computer(network_policy={"mode": "off"})
try:
    print(computer.network_policy)
finally:
    computer.delete()

import { Computer } from "@celestoai/sdk";


const computer = await Computer.create({ networkPolicy: { mode: "off" } });

try {
  console.log(computer.networkPolicy);
} finally {
  await computer.delete();
}

Both examples print the following policy:

{ "mode": "off" }

Use an offline computer with local home storage

An offline computer uses the default home storage. A persistent home is a saved home folder that uses external storage, and it needs internet access. Keep persistent_home=False in Python or persistentHome: false in TypeScript when you select off.

network_policy={"mode": "off"} cannot be combined with persistent_home=True. The TypeScript equivalents, networkPolicy: { mode: "off" } and persistentHome: true, cannot be combined either. The SDK validates this before it creates the computer.

Use the default open setting when your workflow needs a persistent home. Learn more about saved work in Keep work between sessions.

Use Network Control with OpenAI Agents

CelestoSandboxClientOptions accepts the same policy when it creates a hosted computer for an OpenAI agent. Install the OpenAI Agents integration first with pip install "celesto[openai-agents]".

network_controlled_agent.py

import asyncio

from celesto.integrations.openai_agents import (
    CelestoSandboxClient,
    CelestoSandboxClientOptions,
)


async def main() -> None:
    client = CelestoSandboxClient()
    session = await client.create(
        options=CelestoSandboxClientOptions(
            network_policy={"mode": "off"},
        ),
    )
    try:
        print(session.state.computer_id)
    finally:
        await client.delete(session)


asyncio.run(main())

When you reuse a computer with computer_id, the requested policy must match that computer’s saved setting. If it does not, the session fails closed: client.create() raises RuntimeError and does not start the computer, so an agent that expected an offline sandbox never runs against an online one. Omit network_policy when you want to accept whatever setting the existing computer was created with. For a complete agent workflow, see Sandbox an OpenAI agent with Celesto or SmolVM.

Start Here

Features

Guides

Agents and Integrations

Control internet access for AI agent sandboxes

Choose an internet setting

Create an offline computer

Keep the setting with the computer

Use an offline computer with local home storage

Use Network Control with OpenAI Agents

​Choose an internet setting

​Create an offline computer

​Keep the setting with the computer

​Use an offline computer with local home storage

​Use Network Control with OpenAI Agents

​Related pages

Choose an internet setting

Create an offline computer

Keep the setting with the computer

Use an offline computer with local home storage

Use Network Control with OpenAI Agents

Related pages