Anthropic SDK

The Anthropic SDK supports tool use through the Messages API. This guide shows how to define a tool that executes code in Vercel Sandbox, so model-generated code runs in an isolated microVM instead of on your host. Model requests route through AI Gateway, and one set of Vercel credentials covers both the model and the sandbox.

1. Create a new project

   Create a new directory for your project and initialize it:

   terminal

   mkdir anthropic-sandbox
   cd anthropic-sandbox
   npm init -y

2. Install dependencies

   Install the Anthropic SDK and the Sandbox SDK:

   pnpmyarnnpmbun

   Terminal

   pnpm i @anthropic-ai/sdk @vercel/sandbox

   Terminal

   yarn add @anthropic-ai/sdk @vercel/sandbox

   Terminal

   npm i @anthropic-ai/sdk @vercel/sandbox

   Terminal

   bun add @anthropic-ai/sdk @vercel/sandbox

3. Configure authentication

   Link the project to Vercel and pull your development environment. This creates a .env.local file with a VERCEL_OIDC_TOKEN, which authenticates both the Sandbox SDK and AI Gateway:

   terminal

   vercel link
   vercel env pull

   You can also use an AI Gateway API key by setting AI_GATEWAY_API_KEY. Outside Vercel and local development, authenticate the Sandbox SDK with an access token by setting VERCEL_TOKEN, VERCEL_TEAM_ID, and VERCEL_PROJECT_ID.

4. Create your application

   Create a new file called index.ts with the following code:

   index.ts

   import Anthropic from '@anthropic-ai/sdk';
   import { Sandbox } from '@vercel/sandbox';
    
   const gatewayApiKey = process.env.AI_GATEWAY_API_KEY;
   const anthropic = new Anthropic({
     ...(gatewayApiKey
       ? { apiKey: gatewayApiKey }
       : { authToken: process.env.VERCEL_OIDC_TOKEN }),
     baseURL: 'https://ai-gateway.vercel.sh',
   });
    
   const sandbox = await Sandbox.create({ timeout: 300_000 });
    
   try {
     const tools: Anthropic.Tool[] = [
       {
         name: 'run_code',
         description:
           'Run Node.js code in an isolated Vercel Sandbox and return its output.',
         input_schema: {
           type: 'object',
           properties: {
             code: { type: 'string', description: 'Node.js code to execute' },
           },
           required: ['code'],
         },
       },
     ];
    
     const messages: Anthropic.MessageParam[] = [
       {
         role: 'user',
         content:
           "Compute the SHA-256 hex digest of the exact string 'vercel sandbox' with Node.js, then report the digest.",
       },
     ];
    
     let response = await anthropic.messages.create({
       model: 'anthropic/claude-opus-5',
       max_tokens: 16000,
       tools,
       messages,
     });
    
     while (response.stop_reason === 'tool_use') {
       messages.push({ role: 'assistant', content: response.content });
       const toolResults: Anthropic.ToolResultBlockParam[] = [];
       for (const block of response.content) {
         if (block.type !== 'tool_use') continue;
         const { code } = block.input as { code: string };
         const result = await sandbox.runCommand('node', ['-e', code]);
         const output =
           result.exitCode === 0
             ? await result.stdout()
             : `Command failed: ${await result.stderr()}`;
         toolResults.push({
           type: 'tool_result',
           tool_use_id: block.id,
           content: output,
         });
       }
       messages.push({ role: 'user', content: toolResults });
       response = await anthropic.messages.create({
         model: 'anthropic/claude-opus-5',
         max_tokens: 16000,
         tools,
         messages,
       });
     }
    
     for (const block of response.content) {
       if (block.type === 'text') console.log(block.text);
     }
   } finally {
     await sandbox.stop();
   }

   The timeout value gives the sandbox up to five minutes to finish the tool loop. The finally block stops it sooner when the loop finishes or throws.

   The following code:

   • Points the Anthropic SDK at AI Gateway, authenticated with your Vercel credentials
   • Creates a sandbox that is reused across tool calls
   • Declares a run_code tool the model can call
   • Executes each tool call with sandbox.runCommand() and returns the output as a tool_result block
   • Loops while stop_reason is tool_use, then prints the final answer

5. Run the application

   Run your application with the environment variables from .env.local:

   terminal

   npx tsx --env-file=.env.local index.ts

   Claude writes Node.js code, the sandbox executes it, and the final response reports the digest.