Anthropic / Claude

Project sandbox tools into the Anthropic SDK's runnable-tool shape. One array works with both the toolRunner auto-loop and a manual messages.create loop. Use the ai() plugin with the anthropic() framework, or toAnthropicTools().

Installation

@anthropic-ai/sdk is an optional peer dependency.

pnpm add sbox-sdk @anthropic-ai/sdk

Usage

Pass the anthropic() framework to the ai() plugin and sandbox.tools is an array of runnable Claude tools. The simplest path is the beta toolRunner, which runs the whole tool loop for you:

import { createSandboxClient } from "sbox-sdk";
import { e2b } from "sbox-sdk/e2b";
import { ai } from "sbox-sdk/ai";
import { anthropic } from "sbox-sdk/anthropic";
import Anthropic from "@anthropic-ai/sdk";

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

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

const message = await claude.beta.messages.toolRunner({
  model: "claude-sonnet-4-5",
  max_tokens: 1024,
  messages: [
    { role: "user", content: "Create app.js that prints hello, then run it." },
  ],
  tools: sandbox.tools,
});

One array, both loops

Each tool is a runnable tool — it is both a tool definition and an executor — so the same sandbox.tools array also works with a manual messages.create loop:

const res = await claude.beta.messages.create({
  model,
  max_tokens,
  messages,
  tools: sandbox.tools,
});

for (const block of res.content) {
  if (block.type === "tool_use") {
    const tool = sandbox.tools.find((t) => t.name === block.name)!;
    const output = await tool.run(tool.parse(block.input));
    // append a tool_result block with `output`, then continue the conversation
  }
}

Standalone

import { toAnthropicTools } from "sbox-sdk/anthropic";

const tools = toAnthropicTools(sandbox, {
  only: ["sbox_exec", "sbox_fs_read"],
});

Approval

Claude has no native human-in-the-loop, so this adapter enforces the policy inside each tool's run: a deny decision short-circuits, and ask awaits policy.onApprovalRequest when configured. Both return a message Claude reads.

plugins: [
  ai({
    framework: anthropic(),
    policy: {
      defaults: { destructive: "ask" },
      onApprovalRequest: async (req) => askUser(req.title),
    },
  }),
];

Notes

Peer range: @anthropic-ai/sdk@>=0.40. The runnable tools use the SDK's beta tool helpers.