Understanding Sandboxes

Vercel Sandboxes provide on-demand, isolated compute environments for running untrusted code, testing applications, executing AI-generated scripts, and more. Sandboxes are persistent by default: when a sandbox stops, the SDK automatically snapshots its filesystem, and the sandbox configuration is preserved across sessions, so both are restored the next time you resume.

What is a sandbox?

A sandbox is an isolated Linux environment that you create programmatically with the SDK or CLI. Think of it as a secure virtual machine that:

Boots from a Vercel Managed Image, a custom VCR image, or a saved snapshot
Uses Ubuntu, Arch Linux, or any other Linux distribution you need
Has network access for installing packages and making API calls
Automatically stops after a configurable timeout
Provides full root access to install any package or binary

Each sandbox includes configurable isolation:

Filesystem access: A dedicated private filesystem. With persistence enabled (the default), the filesystem is automatically snapshotted on stop and restored when the sandbox resumes. The sandbox configuration is preserved across sessions and reapplied on resume.
Process isolation: Kernel-level isolation ensures code cannot see or access processes in other sandboxes.
Network isolation: Each sandbox has its own network namespace with controlled outbound access.

Sandboxes vs containers

Unlike Docker containers, each sandbox runs in its own Firecracker microVM with a dedicated kernel. This provides stronger isolation than container-based solutions, which makes sandboxes ideal for running untrusted code.

Aspect	Docker containers	Vercel Sandboxes
Isolation	Shares host kernel; relies on namespaces and cgroups	Dedicated kernel per sandbox; full VM isolation
Security	Suitable for trusted code; container escapes are possible	Designed for untrusted code; microVM boundary prevents escapes
Startup time	Sub-second	Milliseconds (Firecracker optimized for fast boot)
Use case	Packaging and deploying applications	Running arbitrary, untrusted code safely

If you already use Docker images to define your environment, store the image in Vercel Container Registry (VCR) and create the sandbox with a custom image. You can also install packages with your system's package manager, or take a snapshot after setup when a Docker image is not needed.

How sandboxes work

When you call Sandbox.create(), Vercel provisions a Firecracker microVM on its infrastructure. This microVM boots an Ubuntu 26.04 image, a custom image from VCR, or a saved snapshot.

The sandbox runs on Vercel's infrastructure, so you don't need to manage servers, scale capacity, or worry about availability. Sandboxes provision in the iad1 region by default. You can choose a different region per sandbox or set a project default.

Here's what happens during the lifecycle:

Provisioning: Vercel allocates compute resources and boots the microVM. Resuming from a snapshot is even faster than starting a fresh sandbox.
Running: Your code executes inside a session. A session is a single running VM instance inside a sandbox. You can run commands, install packages, start servers, and interact with the filesystem.
Stopping: When the timeout expires or you call stop(), the session shuts down. For persistent sandboxes (the default), the SDK automatically snapshots the filesystem so the next session resumes from the same state. For non-persistent sandboxes, the filesystem is discarded.
Resuming: The next SDK call (such as runCommand or writeFiles) on a stopped persistent sandbox starts a new session from the most recent snapshot. You don't need to resume manually.

For long-term storage of data that doesn't belong in the filesystem, write it to external services like databases or object storage.

Sandbox lifecycle

Creating a sandbox

When you're ready to use a sandbox, you can either create a new one from scratch or use a saved snapshot of a sandbox you created previously. Using a snapshot is much faster than creating from scratch because it avoids reinstalling dependencies and repeating setup steps.

Think of it like the difference between booting a fresh OS install versus resuming from a saved state. A new sandbox gives you a clean slate; a snapshot gives you a pre-configured environment ready to go.

Sandboxes are identified by a name that is unique within your project. If you don't provide one, the SDK generates a name for you. Use the same name later with Sandbox.get() or Sandbox.getOrCreate() to resume the sandbox.

To create a sandbox, you can use the CLI, the JS SDK, or the Python SDK:

import { Sandbox } from '@vercel/sandbox';

// Create a new sandbox
const sandbox = await Sandbox.create({ name: 'my-sandbox' });

// Or create from a snapshot
const sandboxFromSnapshot = await Sandbox.create({
  source: { type: 'snapshot', snapshotId: 'snap_abc123' },
});

// Or retrieve an existing sandbox by name (resumes if stopped)
const existing = await Sandbox.get({ name: 'my-sandbox' });

Running commands

Once created, you can run commands inside the sandbox. Commands can run in blocking mode (wait for completion) or detached mode (return immediately).

// Blocking: waits for the command to finish
const result = await sandbox.runCommand('npm', ['install']);
console.log(result.exitCode);

// Detached: returns immediately, useful for servers
const cmd = await sandbox.runCommand({
  cmd: 'npm',
  args: ['run', 'dev'],
  detached: true,
});

// Stream logs from a detached command
for await (const log of cmd.logs()) {
  console.log(log.data);
}

Stopping a sandbox

Sandboxes automatically stop after a timeout. The default timeout is 5 minutes, and the maximum applies to each session, not to the sandbox itself. A sandbox spans as many sessions as you resume it for.

Alternatively, you can stop them manually. stop() resolves once the VM is fully stopped, and returns the final session state. For persistent sandboxes, the resolved value also includes metadata for the snapshot captured during shutdown.

await sandbox.stop();

You can also stop sandboxes from the Vercel Dashboard by navigating to Observability > Sandboxes and clicking Stop Sandbox.

Taking snapshots

Snapshots save the current state of a sandbox, including all installed packages and files. Use snapshots to skip setup time on subsequent runs, checkpoint long-running tasks, or share environments with teammates.

Common use cases

Vercel Sandboxes are ideal for features that require secure, on-demand code execution:

Pattern	Why use sandboxes?	Example
AI code interpreter	LLM-generated code can be unpredictable. Sandboxes ensure it runs in isolation.	An AI assistant that solves math problems by writing and running Python scripts.
Clean test environments	Start fresh for every test run to avoid "works on my machine" issues.	Running unit tests against a clean OS for every commit.
Reproducible infrastructure	Share identical snapshots of environments across teams.	A QA team spinning up an exact replica of a customer's environment.
Temporary debugging	Spin up a throwaway environment to inspect issues without risk.	Investigating a production issue by replicating the environment.

When not to use sandboxes

Sandboxes are not designed to run continuously. They are not suitable for:

Permanent hosting: If you need a server that stays up 24/7, use a traditional VM or Vercel Functions.
Long-term storage of large datasets: The filesystem persists between sessions of a persistent sandbox but it isn't a substitute for a database or object store. Push large or shared data to external services.

Security model

Vercel Sandboxes are designed for running untrusted code safely.

Isolation architecture

Each sandbox runs in its own Firecracker microVM with a dedicated kernel, so you can run processes that require system-level privileges without affecting other sandboxes or the host. These workloads run with sudo and are isolated to your sandbox by the microVM boundary.

Supported workloads include:

Container runtimes: Run Docker and other container engines inside the sandbox to build images or run containerized workloads.
VPN clients: Connect to a VPN provider to reach private networks during a session.
FUSE filesystems: Mount Filesystem in Userspace (FUSE) drivers to attach object storage, network filesystems, or other custom mounts.

These processes require elevated privileges, so run them with sudo. For example, to run a command with elevated privileges through the CLI:

sandbox exec --sudo <name> -- <command>

Outbound network access from these workloads still follows the sandbox firewall network policy. Restrict reachable destinations with a network policy when you run untrusted code.

Resource limits

Every sandbox comes with:

A dedicated private filesystem
Network namespace isolation
Kernel-level process isolation
Strict CPU, memory, and disk limits
Automatic timeouts to prevent runaway processes

These limits prevent resource exhaustion and ensure fair usage across all sandboxes.

Network access

Sandboxes can make outbound HTTP requests by default, so you can install packages from public registries like npm or PyPI. Exposed ports are accessible via a public URL, so be mindful of what services you run.

Internet access from the sandbox can be restricted through network policies defined by the users, as part of the sandbox firewall.

To send a sandbox's public-internet traffic through a Secure Compute network's static IPs, and to reach your own AWS VPC over VPC peering, see Using Secure Compute with Sandbox.

Data privacy

Sandboxes run on Vercel's secure infrastructure, which maintains SOC 2 Type II certification. Since sandboxes are ephemeral, they do not persist data long-term. For specific data residency requirements, consult your plan details or compliance team.