Tool Calling

As covered under Foundations, tools are objects that can be called by the model to perform a specific task. AI SDK Core tools contain three elements:

• description: An optional description of the tool that can influence when the tool is picked.
• parameters: A Zod schema or a JSON schema that defines the parameters. The schema is consumed by the LLM, and also used to validate the LLM tool calls.
• execute: An optional async function that is called with the arguments from the tool call. It produces a value of type RESULT (generic type). It is optional because you might want to forward tool calls to the client or to a queue instead of executing them in the same process.

The tools parameter of generateText and streamText is an object that has the tool names as keys and the tools as values:

import { z } from 'zod';
import { generateText, tool } from 'ai';

const result = await generateText({
  model: yourModel,
  tools: {
    weather: tool({
      description: 'Get the weather in a location',
      parameters: z.object({
        location: z.string().describe('The location to get the weather for'),
      }),
      execute: async ({ location }) => ({
        location,
        temperature: 72 + Math.floor(Math.random() * 21) - 10,
      }),
    }),
  },
  prompt: 'What is the weather in San Francisco?',
});

Tool calling is not restricted to only text generation. You can also use it to render user interfaces (Generative UI).

Multi-Step Calls

Large language models need to know the tool results before they can continue to generate text. This requires sending the tool results back to the model. You can enable this feature by setting the maxSteps setting to a number greater than 1.

When maxSteps is set to a number greater than 1, the language model will be called in a loop when there are tool calls and for every tool call there is a tool result, until there are no further tool calls or the maximum number of tool steps is reached.

Example

In the following example, there are two steps:

1. Step 1
   1. The prompt 'What is the weather in San Francisco?' is sent to the model.
   2. The model generates a tool call.
   3. The tool call is executed.
2. Step 2
   1. The tool result is sent to the model.
   2. The model generates a response considering the tool result.

import { z } from 'zod';
import { generateText, tool } from 'ai';

const { text, steps } = await generateText({
  model: yourModel,
  tools: {
    weather: tool({
      description: 'Get the weather in a location',
      parameters: z.object({
        location: z.string().describe('The location to get the weather for'),
      }),
      execute: async ({ location }) => ({
        location,
        temperature: 72 + Math.floor(Math.random() * 21) - 10,
      }),
    }),
  },
  maxSteps: 5, // allow up to 5 steps
  prompt: 'What is the weather in San Francisco?',
});

Steps

To access intermediate tool calls and results, you can use the steps property in the result object or the streamText onFinish callback. It contains all the text, tool calls, tool results, and more from each step.

Example: Extract tool results from all steps

import { generateText } from 'ai';

const { steps } = await generateText({
  model: openai('gpt-4-turbo'),
  maxSteps: 10,
  // ...
});

// extract all tool calls from the steps:
const allToolCalls = steps.flatMap(step => step.toolCalls);

onStepFinish callback

When using generateText or streamText, you can provide an onStepFinish callback that is triggered when a step is finished, i.e. all text deltas, tool calls, and tool results for the step are available. When you have multiple steps, the callback is triggered for each step.

import { generateText } from 'ai';

const result = await generateText({
  // ...
  onStepFinish({ text, toolCalls, toolResults, finishReason, usage }) {
    // your own logic, e.g. for saving the chat history or recording usage
  },
});

Response Messages

Adding the generated assistant and tool messages to your conversation history is a common task, especially if you are using multi-step tool calls.

Both generateText and streamText have a responseMessages property that you can use to add the assistant and tool messages to your conversation history. It is also available in the onFinish callback of streamText.

The responseMessages property contains an array of CoreMessage objects that you can add to your conversation history:

import { generateText } from 'ai';

const messages: CoreMessage[] = [
  // ...
];

const { responseMessages } = await generateText({
  // ...
  messages,
});

// add the response messages to your conversation history:
messages.push(...responseMessages); // streamText: ...(await responseMessages)

Tool Choice

You can use the toolChoice setting to influence when a tool is selected. It supports the following settings:

• auto (default): the model can choose whether and which tools to call.
• required: the model must call a tool. It can choose which tool to call.
• none: the model must not call tools
• { type: 'tool', toolName: string (typed) }: the model must call the specified tool

import { z } from 'zod';
import { generateText, tool } from 'ai';

const result = await generateText({
  model: yourModel,
  tools: {
    weather: tool({
      description: 'Get the weather in a location',
      parameters: z.object({
        location: z.string().describe('The location to get the weather for'),
      }),
      execute: async ({ location }) => ({
        location,
        temperature: 72 + Math.floor(Math.random() * 21) - 10,
      }),
    }),
  },
  toolChoice: 'required', // force the model to call a tool
  prompt: 'What is the weather in San Francisco?',
});

Tool Execution Options

When tools are called, they receive additional options as a second parameter.

Messages

The messages that were sent to the language model to initiate the response that contained the tool call are forwarded to the tool execution. You can access them in the second parameter of the execute function. In multi-step calls, the messages contain the text, tool calls, and tool results from all previous steps.

import { generateText, tool } from 'ai';

const result = await generateText({
  // ...
  tools: {
    myTool: tool({
      // ...
      execute: async (args, { messages }) => {
        // use the message history in e.g. calls to other language models
        return something;
      },
    }),
  },
});

Abort Signals

The abort signals from generateText and streamText are forwarded to the tool execution. You can access them in the second parameter of the execute function and e.g. abort long-running computations or forward them to fetch calls inside tools.

import { z } from 'zod';
import { generateText, tool } from 'ai';

const result = await generateText({
  model: yourModel,
  abortSignal: myAbortSignal, // signal that will be forwarded to tools
  tools: {
    weather: tool({
      description: 'Get the weather in a location',
      parameters: z.object({ location: z.string() }),
      execute: async ({ location }, { abortSignal }) => {
        return fetch(
          `https://api.weatherapi.com/v1/current.json?q=${location}`,
          { signal: abortSignal }, // forward the abort signal to fetch
        );
      },
    }),
  },
  prompt: 'What is the weather in San Francisco?',
});

Types

Modularizing your code often requires defining types to ensure type safety and reusability. To enable this, the AI SDK provides several helper types for tools, tool calls, and tool results.

You can use them to strongly type your variables, function parameters, and return types in parts of the code that are not directly related to streamText or generateText.

Each tool call is typed with CoreToolCall<NAME extends string, ARGS>, depending on the tool that has been invoked. Similarly, the tool results are typed with CoreToolResult<NAME extends string, ARGS, RESULT>.

The tools in streamText and generateText are defined as a Record<string, CoreTool>. The type inference helpers CoreToolCallUnion<TOOLS extends Record<string, CoreTool>> and CoreToolResultUnion<TOOLS extends Record<string, CoreTool>> can be used to extract the tool call and tool result types from the tools.

import { openai } from '@ai-sdk/openai';
import { CoreToolCallUnion, CoreToolResultUnion, generateText, tool } from 'ai';
import { z } from 'zod';

const myToolSet = {
  firstTool: tool({
    description: 'Greets the user',
    parameters: z.object({ name: z.string() }),
    execute: async ({ name }) => `Hello, ${name}!`,
  }),
  secondTool: tool({
    description: 'Tells the user their age',
    parameters: z.object({ age: z.number() }),
    execute: async ({ age }) => `You are ${age} years old!`,
  }),
};

type MyToolCall = CoreToolCallUnion<typeof myToolSet>;
type MyToolResult = CoreToolResultUnion<typeof myToolSet>;

async function generateSomething(prompt: string): Promise<{
  text: string;
  toolCalls: Array<MyToolCall>; // typed tool calls
  toolResults: Array<MyToolResult>; // typed tool results
}> {
  return generateText({
    model: openai('gpt-4o'),
    tools: myToolSet,
    prompt,
  });
}

Active Tools

Language models can only handle a limited number of tools at a time, depending on the model. To allow for static typing using a large number of tools and limiting the available tools to the model at the same time, the AI SDK provides the experimental_activeTools property.

It is an array of tool names that are currently active. By default, the value is undefined and all tools are active.

import { openai } from '@ai-sdk/openai';
import { generateText } from 'ai';

const { text } = await generateText({
  model: openai('gpt-4o'),
  tools: myToolSet,
  experimental_activeTools: ['firstTool'],
});

Multi-modal Tool Results

In order to send multi-modal tool results, e.g. screenshots, back to the model, they need to be converted into a specific format.

AI SDK Core tools have an optional experimental_toToolResultContent function that converts the tool result into a content part.

Here is an example for converting a screenshot into a content part:

const result = await generateText({
  model: anthropic('claude-3-5-sonnet-20241022'),
  tools: {
    computer: anthropic.tools.computer_20241022({
      // ...
      async execute({ action, coordinate, text }) {
        switch (action) {
          case 'screenshot': {
            return {
              type: 'image',
              data: fs
                .readFileSync('./data/screenshot-editor.png')
                .toString('base64'),
            };
          }
          default: {
            return `executed ${action}`;
          }
        }
      },

      // map to tool result content for LLM consumption:
      experimental_toToolResultContent(result) {
        return typeof result === 'string'
          ? [{ type: 'text', text: result }]
          : [{ type: 'image', data: result.data, mimeType: 'image/png' }];
      },
    }),
  },
  // ...
});

Examples

You can see tools in action using various frameworks in the following examples: