Show HN: "Code Mode" for Vercel AI SDK

2 hours ago 2

Plug-n-play "code mode" tool call scripting for Vercel AI SDK

Inspired by Cloudflare's Code Mode - LLMs are better at writing JavaScript than using synthetic tool calling syntax.

npm install ai tool-scripting

import { z } from 'zod'; import { generateText, tool, stepCountIs } from 'ai'; import { toolScript } from 'tool-scripting'; const tools = { getUserLocation: tool({ description: 'Get user current location', inputSchema: z.object({}), outputSchema: z.string(), // optional outputSchema to help the LLM compose tool calls execute: async () => 'San Francisco, CA', }), getWeather: tool({ description: 'Get weather for a location', inputSchema: z.object({ location: z.string(), }), outputSchema: z.object({ temperature: z.number(), condition: z.string(), }), execute: async ({ location }) => { return { location, temperature: 65, condition: 'foggy' }; }, }), }; // Just wrap your existing generateText (or streamText) const result = await toolScripting(generateText)({ model: 'openai/gpt-5', tools, messages: [ { role: 'assistant', content: 'How can I help?' }, { role: 'user', content: 'Check the weather near me' }, ], stopWhen: stepCountIs(5), });

Converts your tool definitions to a tool call SDK
LLM Generates JavaScript code instead of tool calls
Executes code in secure sandbox (v8 isolate) with tool bindings
Returns whatever the generated code returns

Tool Scripting > Tool Calls

🧠 Better - LLMs excel at JavaScript vs synthetic tool syntax
🔧 Composable - Logic and conditionals between tool calls
🔒 Secure - Sandboxed execution with controlled bindings
🎯 Simple - Just wrap your existing Vercel AI SDK calls

Here's what a traditional series of tool calls looks like (without Tool Scripting):

role: user text: Check the weather near me -- role: assistant type: tool-call toolName: getUserLocation -- role: tool type: tool-result output: San Francisco, CA -- role: assistant type: tool-call toolName: getWeather input: location: San Francisco, CA -- role: tool type: tool-result output: temperature: 65 condition: foggy -- role: assistant text: The weather in San Francisco, CA today is foggy with a temperature of 65°F.

Now, here's the same process with Tool Scripting:

role: user text: Check the weather near me -- role: assistant type: tool-call toolName: runToolScript input: script: const location = await getUserLocation();\nconst weather = await getWeather({ location });\nreturn { location, weather }; -- role: tool type: tool-result output: location: San Francisco, CA weather: temperature: 65 condition: foggy -- role: assistant text: The weather in San Francisco, CA today is foggy with a temperature of 65°F.

💥 In a single LLM step, we composed two tools to get the user's location and then the weather for that location.