Skip to main content
Tool calling lets models invoke deterministic functions for things they’re bad at: counting, lookups, validation, and calling external systems. It makes responses more reliable and lets you combine LLM reasoning with real-world actions.

​
Local

By default, the LLM only suggests a tool call; the actual execution runs in your own agent code and returns the result.
Vercel AI SDK
import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
import { generateText, tool } from "ai";
import { z } from "zod";

const hebo = createOpenAICompatible({
  apiKey: process.env.HEBO_API_KEY,
  baseURL: "https://gateway.hebo.ai/v1",
});

const { text } = await generateText({
  model: hebo("openai/gpt-oss-20b"),
  prompt: "How many r's are in Strawberry?",
  tools: {
    countLetters: tool({
      description: "Count letter occurrences in a word",
      parameters: z.object({
        word: z.string(),
        letters: z.string(),
      }),
      execute: async ({ word, letters }) => {
        const count = [...word].filter((c) => letters.includes(c)).length;
        return { count };
      },
    }),
  },
});

console.log(text);

​
MCP

MCP (Model Context Protocol) is a standard way to expose tools over a remote server. For read-to-use MCP tools, see the our MCP AIKit docs.