Overview

Give your agent a sandbox

The AI layer turns a Sandbox into tools an LLM can call — run commands, read and write files, execute code, expose ports — shaped for whichever agent framework you use, with a single approval policy you control.

You write the integration once. Pick an AI-provider plugin, and the sandbox grows a correctly-typed sandbox.tools you hand straight to your framework:

import { createSandboxClient } from "sbox-sdk";
import { e2b } from "sbox-sdk/e2b";
import { ai } from "sbox-sdk/ai";
import { aiSdk } from "sbox-sdk/ai-sdk";
import { generateText } from "ai";

const client = createSandboxClient({
  provider: e2b({ apiKey: process.env.E2B_API_KEY! }),
  plugins: [ai({ framework: aiSdk() })],
});

const sandbox = await client.create({ template: "node" });

// `sandbox.tools` is typed for the AI SDK and already filtered to what E2B supports.
const result = await generateText({
  model,
  prompt: "Create app.js that prints hello, then run it.",
  tools: sandbox.tools,
});

The framework you pass to ai() decides the shape of sandbox.tools — a tool map for the AI SDK, an array for OpenAI Agents, runnable tools for Claude. Swap e2b() for vercel() and the toolset re-shapes itself to the new provider's capabilities, with zero changes to your agent code.

How it fits together

One provider-neutral tool registry is projected into every framework. You never hand-write the same tools per framework:

   Sandbox ──▶ agent-tools (10 capability-gated tools + policy)
                     │  projection (shape only)
        ┌────────────┼─────────────┬──────────────┬───────────────┐
        ▼            ▼             ▼              ▼               ▼
     ai-sdk       mastra        openai        anthropic       langchain
   tool() map  createTool map  tool() array  runnable[]     tool() array

• AI providers — install an AI-provider plugin on the client and sandbox.tools is ready to use, typed for your framework.
• Tools — the 10 canonical tools, each mapped to the core Sandbox API and gated by the provider's capabilities.
• Approval — one SandboxPolicy decides allow / ask / deny per call; each framework maps ask to its native human-in-the-loop.

Two ways to use it

Every framework subpath ships both layers:

1. Plugin (batteries included) — plugins: [ai({ framework: aiSdk() })] → sandbox.tools. The shape follows the framework you pass.
2. Standalone function (composable) — toAISDKTools(sandbox) returns the same tools explicitly, for when you want to filter, merge, or build them yourself.

import { toAISDKTools } from "sbox-sdk/ai-sdk";

const tools = toAISDKTools(sandbox, {
  policy: { defaults: { destructive: "ask" } },
});

Supported frameworks

All used as ai({ framework: … }):

Each framework SDK is an optional peer dependency — install only the one you use.