Adding AI Tools

Overview

AI tools are functions that the AI can call to perform actions or retrieve information. Gorkie uses the Vercel AI SDK’s tool() function to define tools with type-safe parameters and execution logic. There are two patterns for creating tools:

Context-aware tools - Need access to Slack context and streaming (most common)
Stateless tools - Pure functions that don’t need context

Tool Anatomy

Every tool has:

Description - Tells the AI when to use the tool
Input schema - Zod schema defining parameters
Execute function - The actual implementation
Optional lifecycle hooks - onInputStart, onInputComplete

Pattern 1: Context-Aware Tools

Use this pattern when your tool needs:

Access to Slack client (to send messages, reactions, etc.)
Access to the message context (channel, user, timestamp)
Ability to stream status updates

Factory Pattern

Context-aware tools use a factory function:

import { tool } from 'ai';
import { z } from 'zod';
import { createTask, finishTask, updateTask } from '~/lib/ai/utils/task';
import logger from '~/lib/logger';
import type { SlackMessageContext, Stream } from '~/types';

export const myTool = ({
  context,
  stream,
}: {
  context: SlackMessageContext;
  stream: Stream;
}) =>
  tool({
    description: 'What this tool does and when to use it',
    inputSchema: z.object({
      param1: z.string().describe('Description for the AI'),
      param2: z.number().optional().describe('Optional parameter'),
    }),
    onInputStart: async ({ toolCallId }) => {
      // Optional: Create a task for status tracking
      await createTask(stream, {
        taskId: toolCallId,
        title: 'Starting operation',
        status: 'pending',
      });
    },
    execute: async ({ param1, param2 }, { toolCallId }) => {
      // Update task status
      const task = await updateTask(stream, {
        taskId: toolCallId,
        title: 'Performing operation',
        details: param1,
        status: 'in_progress',
      });

      try {
        // Your tool logic here
        const result = await performOperation(param1, param2);

        // Log success
        logger.info({ param1, param2 }, 'Operation completed');

        // Mark task complete
        await finishTask(stream, {
          status: 'complete',
          taskId: task,
          output: 'Operation successful',
        });

        // Return result to AI
        return {
          success: true,
          data: result,
        };
      } catch (error) {
        // Log error
        logger.error({ error, param1 }, 'Operation failed');

        // Mark task failed
        await finishTask(stream, {
          status: 'error',
          taskId: task,
          output: 'Operation failed',
        });

        // Return error to AI
        return {
          success: false,
          error: error instanceof Error ? error.message : String(error),
        };
      }
    },
  });

Real Example: React Tool

Adds emoji reactions to Slack messages:

server/lib/ai/tools/chat/react.ts

import { tool } from 'ai';
import { z } from 'zod';
import { createTask, finishTask, updateTask } from '~/lib/ai/utils/task';
import logger from '~/lib/logger';
import type { SlackMessageContext, Stream } from '~/types';
import { getContextId } from '~/utils/context';
import { errorMessage, toLogError } from '~/utils/error';

export const react = ({
  context,
  stream,
}: {
  context: SlackMessageContext;
  stream: Stream;
}) =>
  tool({
    description:
      'Add emoji reactions to the current Slack message. Provide emoji names without surrounding colons.',
    inputSchema: z.object({
      emojis: z
        .array(z.string().min(1))
        .nonempty()
        .describe('Emoji names to react with (unicode or custom names).'),
    }),
    onInputStart: async ({ toolCallId }) => {
      await createTask(stream, {
        taskId: toolCallId,
        title: 'Adding reaction',
        status: 'pending',
      });
    },
    execute: async ({ emojis }, { toolCallId }) => {
      const ctxId = getContextId(context);
      const channelId = context.event.channel;
      const messageTs = context.event.ts;

      if (!(channelId && messageTs)) {
        logger.warn(
          { ctxId, channel: channelId, messageTs, emojis },
          'Failed to add Slack reactions: missing channel or message id'
        );
        return { success: false, error: 'Missing Slack channel or message id' };
      }

      const task = await updateTask(stream, {
        taskId: toolCallId,
        title: 'Adding reaction',
        details: emojis.join(', '),
        status: 'in_progress',
      });

      try {
        for (const emoji of emojis) {
          await context.client.reactions.add({
            channel: channelId,
            name: emoji.replace(/:/g, ''),
            timestamp: messageTs,
          });
        }

        logger.info(
          { ctxId, channel: channelId, messageTs, emojis },
          'Added reactions'
        );

        await finishTask(stream, {
          status: 'complete',
          taskId: task,
          output: `Added: ${emojis.join(', ')}`,
        });

        return {
          success: true,
          content: `Added reactions: ${emojis.join(', ')}`,
        };
      } catch (error) {
        logger.error(
          {
            ...toLogError(error),
            ctxId,
            channel: channelId,
            messageTs,
            emojis,
          },
          'Failed to add Slack reactions'
        );

        await finishTask(stream, {
          status: 'error',
          taskId: task,
          output: errorMessage(error),
        });

        return {
          success: false,
          error: errorMessage(error),
        };
      }
    },
  });

Pattern 2: Stateless Tools

Use this pattern for tools that:

Don’t need Slack context
Are pure functions
Call external APIs or perform calculations

Simple Export

Stateless tools are exported directly:

import { tool } from 'ai';
import { z } from 'zod';
import logger from '~/lib/logger';
import type { Stream } from '~/types';

export const myStatelessTool = ({ stream }: { stream: Stream }) =>
  tool({
    description: 'Performs a stateless operation',
    inputSchema: z.object({
      query: z.string().min(1).describe('The search query'),
    }),
    execute: async ({ query }) => {
      try {
        const result = await externalApiCall(query);
        return result;
      } catch (error) {
        logger.error({ error, query }, 'API call failed');
        throw error;
      }
    },
  });

Real Example: Get Weather

Fetches weather data from an external API:

server/lib/ai/tools/chat/get-weather.ts

import { tool } from 'ai';
import { z } from 'zod';
import { createTask, finishTask, updateTask } from '~/lib/ai/utils/task';
import logger from '~/lib/logger';
import type { SlackMessageContext, Stream } from '~/types';
import { toLogError } from '~/utils/error';

export const getWeather = ({
  stream,
}: {
  context: SlackMessageContext;
  stream: Stream;
}) =>
  tool({
    description: 'Get the current weather at a location',
    inputSchema: z.object({
      latitude: z.number(),
      longitude: z.number(),
    }),
    onInputStart: async ({ toolCallId }) => {
      await createTask(stream, {
        taskId: toolCallId,
        title: 'Getting weather',
        status: 'pending',
      });
    },
    execute: async ({ latitude, longitude }, { toolCallId }) => {
      const task = await updateTask(stream, {
        taskId: toolCallId,
        title: 'Getting weather',
        details: `${latitude}, ${longitude}`,
        status: 'in_progress',
      });

      try {
        const response = await fetch(
          `https://api.open-meteo.com/v1/forecast?latitude=${latitude}&longitude=${longitude}&current=temperature_2m&hourly=temperature_2m&daily=sunrise,sunset&timezone=auto`
        );

        if (!response.ok) {
          throw new Error(
            `Weather API request failed with status ${response.status}`
          );
        }

        const weatherData: unknown = await response.json();
        await finishTask(stream, { status: 'complete', taskId: task });
        return weatherData;
      } catch (error) {
        logger.error({ ...toLogError(error) }, 'Error in getWeather');
        await finishTask(stream, {
          status: 'error',
          taskId: task,
          output: 'Failed to fetch weather',
        });
        return {
          success: false,
          error: 'Failed to fetch weather',
        };
      }
    },
  });

Zod Schema Guidelines

Use Descriptive Descriptions

The AI uses parameter descriptions to understand how to use the tool:

inputSchema: z.object({
  query: z
    .string()
    .min(1)
    .max(500)
    .describe(
      "The web search query. Be specific and clear about what you're looking for."
    ),
  numResults: z
    .number()
    .int()
    .min(1)
    .max(10)
    .optional()
    .describe('Number of search results to return (default: 5)'),
}),

Constrain Values

Use Zod’s validation to prevent invalid inputs:

inputSchema: z.object({
  temperature: z.number().min(0).max(2).default(0.7),
  maxTokens: z.number().int().min(1).max(4000).optional(),
  emojis: z.array(z.string().min(1)).nonempty().max(5),
}),

Provide Defaults

inputSchema: z.object({
  type: z.enum(['reply', 'message']).default('reply'),
  offset: z.number().int().min(0).optional().default(0),
}),

Return Value Conventions

Success Response

Return structured data the AI can understand:

return {
  success: true,
  data: result,
  message: 'Operation completed successfully',
};

Error Response

Always include error details:

return {
  success: false,
  error: errorMessage(error),
};

Rich Data

For complex results, return structured objects:

return {
  success: true,
  results: [
    { title: 'Result 1', url: 'https://...', snippet: '...' },
    { title: 'Result 2', url: 'https://...', snippet: '...' },
  ],
  count: 2,
};

Task Tracking

Create Task

Start tracking when the tool begins:

onInputStart: async ({ toolCallId }) => {
  await createTask(stream, {
    taskId: toolCallId,
    title: 'Starting operation',
    status: 'pending',
  });
},

Update Task

Show progress during execution:

const task = await updateTask(stream, {
  taskId: toolCallId,
  title: 'Processing data',
  details: 'Processing 10 items...',
  status: 'in_progress',
});

Finish Task

Mark complete or failed:

// Success
await finishTask(stream, {
  status: 'complete',
  taskId: task,
  output: 'Processed 10 items successfully',
  sources: [{ type: 'url', text: 'Source', url: 'https://...' }],
});

// Error
await finishTask(stream, {
  status: 'error',
  taskId: task,
  output: errorMessage(error),
});

Registering Tools

Add your tool to the orchestrator:

Create the tool file

Create a new file in server/lib/ai/tools/chat/:

touch server/lib/ai/tools/chat/my-new-tool.ts

Import in the orchestrator

Open server/lib/ai/agents/orchestrator.ts and import your tool:

import { myNewTool } from '~/lib/ai/tools/chat/my-new-tool';

Add it to the tools object:

const tools = {
  reply: reply({ context, stream }),
  react: react({ context, stream }),
  myNewTool: myNewTool({ context, stream }),
  // ... other tools
};

Test the tool

Restart the dev server and test in Slack:

bun run dev

Ask the bot to use your new tool!

Example: Creating a New Tool from Scratch

Let’s create a tool that fetches a random programming joke:

server/lib/ai/tools/chat/get-joke.ts

import { tool } from 'ai';
import { z } from 'zod';
import { createTask, finishTask, updateTask } from '~/lib/ai/utils/task';
import logger from '~/lib/logger';
import type { SlackMessageContext, Stream } from '~/types';
import { toLogError } from '~/utils/error';

interface JokeResponse {
  type: string;
  setup?: string;
  delivery?: string;
  joke?: string;
  error?: boolean;
}

export const getJoke = ({
  stream,
}: {
  context: SlackMessageContext;
  stream: Stream;
}) =>
  tool({
    description:
      'Get a random programming joke. Use this when the user asks for humor or a joke.',
    inputSchema: z.object({
      category: z
        .enum(['programming', 'pun', 'any'])
        .optional()
        .default('programming')
        .describe('The category of joke to fetch'),
    }),
    onInputStart: async ({ toolCallId }) => {
      await createTask(stream, {
        taskId: toolCallId,
        title: 'Fetching joke',
        status: 'pending',
      });
    },
    execute: async ({ category }, { toolCallId }) => {
      const task = await updateTask(stream, {
        taskId: toolCallId,
        title: 'Fetching joke',
        details: `Category: ${category}`,
        status: 'in_progress',
      });

      try {
        const response = await fetch(
          `https://v2.jokeapi.dev/joke/${category}?safe-mode`
        );

        if (!response.ok) {
          throw new Error(`Joke API returned status ${response.status}`);
        }

        const data = (await response.json()) as JokeResponse;

        if (data.error) {
          throw new Error('Failed to fetch joke');
        }

        const joke =
          data.type === 'twopart' && data.setup && data.delivery
            ? `${data.setup}\n${data.delivery}`
            : data.joke ?? 'No joke available';

        logger.info({ category, type: data.type }, 'Fetched joke');

        await finishTask(stream, {
          status: 'complete',
          taskId: task,
          output: 'Got a joke!',
        });

        return {
          success: true,
          joke,
          category: data.type,
        };
      } catch (error) {
        logger.error({ ...toLogError(error), category }, 'Failed to fetch joke');

        await finishTask(stream, {
          status: 'error',
          taskId: task,
          output: 'Failed to fetch joke',
        });

        return {
          success: false,
          error: 'Failed to fetch joke',
        };
      }
    },
  });

Now register it in the orchestrator, and the AI can fetch jokes on demand!

Best Practices

Clear Descriptions

Write clear tool descriptions that tell the AI:

What the tool does
When to use it
What parameters it needs

Validate Inputs

Use Zod schemas to validate all inputs:

Constrain string lengths
Validate number ranges
Require specific enum values
Use .nonempty() for arrays that must have items

Handle Errors Gracefully

Always catch and log errors:

Use try-catch blocks
Log errors with context
Return structured error responses
Update task status to ‘error’

Log with Context

Include relevant information in logs:

logger.info(
  { userId, channel, action: 'sendMessage' },
  'Message sent successfully'
);

Stream Updates

Keep users informed with task updates:

Create task on start
Update during execution
Finish with success or error

Return Structured Data

Make it easy for the AI to use your results:

Use consistent response shapes
Include success boolean
Provide meaningful error messages
Return rich data when helpful

Next Steps

Review existing tools for inspiration
Learn about code style guidelines
Explore the Vercel AI SDK documentation

Development Guide

API Reference

Overview

Tool Anatomy

Pattern 1: Context-Aware Tools

Factory Pattern

Real Example: React Tool

Pattern 2: Stateless Tools

Simple Export

Real Example: Get Weather

Zod Schema Guidelines

Use Descriptive Descriptions

Constrain Values

Provide Defaults

Return Value Conventions

Success Response

Error Response

Rich Data

Task Tracking

Create Task

Update Task

Finish Task

Registering Tools

Example: Creating a New Tool from Scratch

Best Practices

Clear Descriptions

Validate Inputs

Handle Errors Gracefully

Log with Context

Stream Updates

Return Structured Data

Next Steps

Build docs developers (and LLMs) love

Development Guide

API Reference

​Overview

​Tool Anatomy

​Pattern 1: Context-Aware Tools

​Factory Pattern

​Real Example: React Tool

​Pattern 2: Stateless Tools

​Simple Export

​Real Example: Get Weather

​Zod Schema Guidelines

​Use Descriptive Descriptions

​Constrain Values

​Provide Defaults

​Return Value Conventions

​Success Response

​Error Response

​Rich Data

​Task Tracking

​Create Task

​Update Task

​Finish Task

​Registering Tools

​Example: Creating a New Tool from Scratch

​Best Practices

​Clear Descriptions

​Validate Inputs

​Handle Errors Gracefully

​Log with Context

​Stream Updates

​Return Structured Data

​Next Steps

Build docs developers (and LLMs) love

Overview

Tool Anatomy

Pattern 1: Context-Aware Tools

Factory Pattern

Real Example: React Tool

Pattern 2: Stateless Tools

Simple Export

Real Example: Get Weather

Zod Schema Guidelines

Use Descriptive Descriptions

Constrain Values

Provide Defaults

Return Value Conventions

Success Response

Error Response

Rich Data

Task Tracking

Create Task

Update Task

Finish Task

Registering Tools

Example: Creating a New Tool from Scratch

Best Practices

Clear Descriptions

Validate Inputs

Handle Errors Gracefully

Log with Context

Stream Updates

Return Structured Data

Next Steps