AI providers

An AI-provider plugin is the first plugin kind. It grafts a framework-shaped sandbox.tools onto every sandbox the client builds. This page covers what's specific to AI providers; for the general plugin mechanism, see Plugins.

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

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

const sandbox = await client.create();
sandbox.tools; // typed for the AI SDK, ready for generateText({ tools })

The shape follows the provider and the framework

sandbox.tools is resolved on two axes at once:

• Capability-gated to the provider — a tool only appears if the provider supports it. On a provider without a code interpreter, sbox_run_code simply isn't there. See Tools.
• Typed for the framework — the framework you pass to ai() decides the type: a tool map for the AI SDK, an array for OpenAI Agents, runnable tools for Claude.

Swap e2b() for vercel() and sandbox.tools re-shapes to the new provider's capabilities with no change to your agent code.

One AI-provider per client

An agent runs in exactly one framework — you hand sandbox.tools to a single generateText(...) or new Agent({ tools }) call — so a single shape is correct. The client throws if you pass two AI-provider plugins:

// ❌ throws: a client supports one AI-provider plugin
createSandboxClient({
  provider: e2b(),
  plugins: [ai({ framework: aiSdk() }), ai({ framework: openaiAgents() })],
});

Need two framework shapes from one sandbox (libraries, benchmarks)? Use the standalone toXTools(sandbox) functions, or read the provider-neutral registry with createSandboxTools(sandbox). Other plugin kinds stack freely alongside an AI provider.

Policy per sandbox

The framework is fixed on the client, but a sandbox's trust posture often isn't — you may run trusted internal work in one sandbox and untrusted user code in another. Pass per-sandbox overrides as the second argument to create:

const client = createSandboxClient({
  provider: e2b(),
  plugins: [ai({ framework: aiSdk(), policy: trustedDefaults })], // default policy
});

// Tighten approval for this sandbox only:
const box = await client.create(spec, {
  policy: { defaults: { mutating: "ask", destructive: "deny" } },
});

Overrides apply to policy, only, and forbid — never the framework. See Approval.

Configuring a plugin

Every AI-provider plugin accepts the same options as its standalone function — a policy, an only allow-list, and a forbid list:

plugins: [
  ai({
    framework: aiSdk(),
    only: ["sbox_exec", "sbox_fs_read", "sbox_fs_write"],
    policy: { forbid: ["sbox_set_egress"] },
  }),
];

See the per-framework pages for each plugin's output shape and approval behavior: Vercel AI SDK, Mastra, OpenAI Agents, Anthropic, LangChain.