Skip to content
Sandbox SDK
Esc
↑↓navigate↵open⌘Jpreview
On this page

AI SDK

Give an AI SDK agent sandbox tools with sandboxTools from @opencoredev/sandbox-sdk/ai. Works with generateText, streamText, and useChat, and streams command output.

sandboxTools(box) from @opencoredev/sandbox-sdk/ai returns an AI SDK tool set. Pass it as tools to generateText or streamText and the model can run commands, read and edit files, and start dev servers in the sandbox.

For the tools themselves and how to limit them, see Give your agent a sandbox.

Install

npm install @opencoredev/sandbox-sdk ai zod
pnpm add @opencoredev/sandbox-sdk ai zod
yarn add @opencoredev/sandbox-sdk ai zod
bun add @opencoredev/sandbox-sdk ai zod

Needs AI SDK 7 and Zod 4.6. Some cloud providers also need their own SDK; Install lists them.

Run an agent

import { generateText, stepCountIs } from "ai";
import { createSandbox } from "@opencoredev/sandbox-sdk";
import { e2b } from "@opencoredev/sandbox-sdk/e2b";
import { sandboxTools } from "@opencoredev/sandbox-sdk/ai";

await using box = await createSandbox(e2b());
await box.git.clone("https://github.com/acme/web.git", { dir: "app", depth: 1 });

const { text } = await generateText({
  model: "anthropic/claude-opus-5-5",
  tools: sandboxTools(box, { root: "/workspace/app" }),
  stopWhen: stepCountIs(30),
  prompt: "Run the tests in /workspace/app and fix the first failure.",
});

console.log(text);

The model string goes through the Vercel AI Gateway, so set AI_GATEWAY_API_KEY on your server. Any AI SDK provider works in its place.

The agent calls run to execute the tests, read_file to look at the failure, and edit_file to change it. You get its final summary in text. stopWhen caps the number of steps, so a confused model can’t loop forever.

Stream to a chat UI

With streamText, return the stream from a route. Don’t use await using here, because the stream outlives the function body. Close the sandbox when the response finishes instead.

import { convertToModelMessages, stepCountIs, streamText, type UIMessage } from "ai";
import { createSandbox } from "@opencoredev/sandbox-sdk";
import { e2b } from "@opencoredev/sandbox-sdk/e2b";
import { sandboxTools } from "@opencoredev/sandbox-sdk/ai";

export async function POST(request: Request) {
  const { messages, chatId }: { messages: UIMessage[]; chatId: string } = await request.json();

  const box = await createSandbox(e2b(), { key: `chat:${chatId}`, timeoutMs: 60 * 60_000 });

  const result = streamText({
    model: "anthropic/claude-opus-5-5",
    messages: await convertToModelMessages(messages),
    tools: sandboxTools(box),
    stopWhen: stepCountIs(30),
    onFinish: () => box.close(),
  });

  return result.toUIMessageStreamResponse();
}

The key gives each chat its own sandbox, so a follow-up message finds the files the last one left. timeoutMs makes E2B delete a forgotten chat’s sandbox after an hour.

Live output

The AI SDK tools are streaming tools. While a command runs, each chunk of output arrives as a preliminary tool result, so a useChat UI shows output as it happens. The model only sees the final, compact result through toModelOutput.

To send full output somewhere else, such as a log panel or your own database, pass onEvent:

import { createSandbox } from "@opencoredev/sandbox-sdk";
import { local } from "@opencoredev/sandbox-sdk/local";
import { sandboxTools } from "@opencoredev/sandbox-sdk/ai";

await using box = await createSandbox(local());

const tools = sandboxTools(box, {
  onEvent: (event) => process.stdout.write(`[${event.tool}] ${event.text}`),
});

Events carry redacted output. Secrets passed to createSandbox never appear.

Cancellation

When the AI SDK aborts a tool call, for example because the user stopped the stream, the tool aborts the command inside the sandbox too.

A full example

Build an AI app builder puts this together with Next.js, Vercel Sandbox, and a live preview. The repository’s examples/coding-agent is a command-line agent that clones a repo, works on a task, and prints the diff.

Was this page helpful?