Install the required dependencies

npm install ai @ai-sdk/openai freestyle-sandboxes

Get your Freestyle API Key from the Freestyle Dashboard

Set up the Code Executor

The simplest code executor looks like this:

import { executeTool } from "freestyle-sandboxes/ai";

const codeExecutor = executeTool({
  apiKey: process.env.FREESTYLE_API_KEY!,
});

You can also pass in any nodeModules, environment variables, timeout, or network restrictions you need.

Here's an example with access to the resend and octokit node modules, and environment variables for RESEND_API_KEY and GITHUB_PERSONAL_ACCESS_TOKEN:

import { executeTool } from "freestyle-sandboxes/ai";

const codeExecutor = executeTool({
  apiKey: process.env.FREESTYLE_API_KEY!,
  nodeModules: {
    resend: "4.0.1",
    octokit: "4.1.0",
  },
  envVars: {
    RESEND_API_KEY: process.env.RESEND_API_KEY!,
    GITHUB_PERSONAL_ACCESS_TOKEN: process.env.GITHUB_PERSONAL_ACCESS_TOKEN!,
  },
});

Set up the Vercel AI SDK Agent

const openai = createOpenAI({
  compatibility: "strict",
  apiKey: process.env.OPENAI_API_KEY!,
});

const { text, steps } = await generateText({
  model: openai("gpt-4o"),
  tools: {
    codeExecutor,
  },
  maxSteps: 5,
  maxRetries: 0,
  prompt: "Ask the AI to do whatever you want",
});

Notes

You can actually give it multiple code executors with different node modules, different environment variables, and different network restrictions.
When the AI writes invalid code, or the code gets errors they will be returned in the response. If the AI has steps left, it will try to fix the code and continue.
If you add a node module we haven't seen before, the first time you run it it will take longer because we have to install the node module. After that, it will be cached and return to normal speed.

Vercel AI SDK

Install the required dependencies

Set up the Code Executor

Set up the Vercel AI SDK Agent

Notes

On this page