Vercel Sandbox

sandbox-vercel provides code sandbox capabilities specifically for the Vercel platform, allowing code generated by AI models to be executed in an isolated container environment.

The Vercel provider runs the agent and its task environment in two separate Sandboxes.

Vercel provides a Python SDK that can call Vercel products such as Vercel Blob, Queues, Workflows, Sandbox, and the AI SDK for Python.

A Drive in Vercel Drives is a persistent volume independent of the Sandbox; the Sandbox associates it with a file system path via mounting, with reads and writes below the mount point landing on the Drive and those outside being temporary storage.

For teams putting sensitive code into third-party sandboxes, data residency and access control policies need to be evaluated.

The current material does not disclose the performance specifications, concurrency limits, or billing method of Vercel Sandbox.

When the sandbox is stopped, the system saves files and conversation history, and after restarting you can continue from the last progress; active running processes will be stopped along with the sandbox.

Vercel explicitly states in the announcement that the change to 2 GB per layer suits images containing large toolchains or dependency sets, with the typical scenario being Vercel Sandbox environments customized for coding agent evaluations.

For teams using Vercel Sandbox to run coding agent evaluations, raising the per-layer limit to 2 GB makes it easier to package customized environments into a single large layer, reducing the assembly work in the environment preparation stage.

The agent's file operations within Vercel Sandbox are by default confined to that runtime environment, and data is lost once the sandbox is destroyed.

Perform the unmount operation before stopping the sandbox; content written by the agent will remain in Box, unaffected even if the sandbox is subsequently deleted.

In the layered architecture of the contract review example, AI Gateway handles model calls, Vercel Sandbox provides the execution environment, and Box Mount serves as the file read/write channel; the three are decoupled, making it easy to replace any component.

Box Mount synchronizes a Box folder with a directory inside a Vercel Sandbox, so the agent's reads and writes to files in the sandbox are reflected in Box in real time.

The Box Mount sync process and the agent run as different sandbox user identities, and Box's authentication credentials are isolated from the agent's workspace.

Typical use cases for custom monitors include AI Gateway costs, Core Web Vitals, failure rates of Workflows and external APIs, and Vercel Sandbox CPU time.

Vercel Sandbox is a temporary compute environment provided by the platform, billed by CPU time.

The Vercel Sandbox TypeScript SDK provides writeStdin() to write data, closeStdin() to end input, and wait() to wait for the process to exit and read stdout.

In the Vercel Sandbox TypeScript SDK, besides a boolean, the stdin parameter can also be bound to a Node.js readable stream (such as process.stdin), and the SDK automatically closes the command's standard input when the stream ends.

The typical stdin use case officially cited by Vercel Sandbox is stdio MCP servers, such as agent-browser.

Message-based protocols such as JSON-RPC carry no framing information over pipes, so the writer must define message boundaries itself (such as line delimiting or length prefixes), and the SDK only forwards bytes.