Show HN: "Code Mode" for Vercel AI SDK (github.com)

Vercel released a plug-and-play "Code Mode" (tool-scripting) layer for the Vercel AI SDK that lets LLMs emit actual JavaScript to orchestrate tools instead of relying on synthetic tool-call JSON. Inspired by Cloudflare’s Code Mode, the package wraps your existing generateText/streamText calls so the model generates JS that calls predefined tool bindings (declared with tool(), including zod input/output schemas and execute hooks). The generated code runs inside a secure V8 isolate with controlled bindings and returns whatever value the script produces—so a single LLM step can compose multiple tools, perform conditionals, and return structured results. This matters because LLMs are empirically better at producing JavaScript than adhering to rigid tool-call protocols; tool scripting reduces brittle multi-step tool-call exchanges, enables richer control flow and error handling on the model side, and simplifies developer integration with the Vercel AI SDK. Technical details: npm package example uses generateText + toolScripting, supports TypeScript/JavaScript, Node.js 18+, and works with models like openai/gpt-5; zod schemas and optional outputSchema help the model compose calls. The runtime enforces sandboxing and limited bindings for security. MIT-licensed, it’s positioned as a composable, secure alternative to classical tool-call orchestration for building multi-tool LLM agents.