Vercel Sandbox adds live process input for stdio MCP tools

Send messages after a command has started

Vercel’s October 26 Sandbox update adds standard‑input support for running commands. An application can send data after starting a process, rather than supplying everything in its initial invocation. Vercel names stdio Model Context Protocol servers, including agent‑browser, as a use case.

For an AI application that talks to a tool server over standard input and output, the useful change is a continuing input channel. The application still has to send the messages that its chosen server expects and handle the responses. The announcement establishes an input mechanism; it is not a claim that every MCP server has been tested.

TypeScript commands need input and detached mode enabled

In the TypeScript example, runCommand() enables both stdin: true and detached: true. The caller then uses writeStdin() to send text, closeStdin() to end input and wait() to wait for the result. The example runs cat and reads back the text it receives. OpenTools has read the example, not executed it.

The SDK also accepts a Node.js readable stream as stdin. Vercel says it forwards the stream’s data and closes the command’s input when that stream ends. That gives applications a stream‑based option as well as explicit writes. The point at which input ends remains part of the process lifecycle.

Python uses communicate to close input and collect output

The Python example creates a process with stdin=subprocess.PIPE and sends bytes with server.stdin.send(). It uses communicate() to close standard input, collect remaining output and wait for the process to exit. These are the Python API names; the notice does not present them as interchangeable with TypeScript’s methods.

For a command receiving all its input at once, Vercel also points to passing input directly to Python’s run_process(). Developers can therefore distinguish a continuing conversation with a running process from a one‑shot input payload.

The changelog does not give a minimum SDK version, new price, regional availability list or new process limit. It also does not describe standard input as an authorization or message‑validation control. Those boundaries need to be checked separately when connecting an agent to a tool.

AI‑generated editorial illustration by OpenTools Team: blank paper entering and leaving an unbranded object is a metaphor for process input and output. This is not Vercel Sandbox’s interface, a technical architecture diagram or a photographed product.

Share this article

Related News

Hugging Face adds a versioned lookup for Trainium model tests

Cohere North adds PitchBook connector for licensed customers with SSO

Copilot metrics now name internal skills, agents and MCP servers

Vercel Sandbox adds live process input for stdio MCP tools

Send messages after a command has started

TypeScript commands need input and detached mode enabled

Python uses communicate to close input and collect output

Sources

Tags

Vercel Sandbox adds live process input for stdio MCP tools

Send messages after a command has started

TypeScript commands need input and detached mode enabled

Python uses communicate to close input and collect output

Sources

Tags