ai-core/tool-calling

# Tool Calling

This skill builds on ai-core. Read it first for critical rules.

## Setup

Complete end-to-end example: shared definition, server tool, client tool, server route, React client.

```typescript
// tools/definitions.ts
import { toolDefinition } from '@tanstack/ai'
import { z } from 'zod'

export const getProductsDef = toolDefinition({
  name: 'get_products',
  description: 'Search for products in the catalog',
  inputSchema: z.object({
    query: z.string().meta({ description: 'Search keyword' }),
    limit: z.number().optional().meta({ description: 'Max results' }),
  }),
  outputSchema: z.object({
    products: z.array(
      z.object({ id: z.string(), name: z.string(), price: z.number() }),
    ),
  }),
})

export const updateCartUIDef = toolDefinition({
  name: 'update_cart_ui',
  description: 'Update the shopping cart UI with item count',
  inputSchema: z.object({ itemCount: z.number(), message: z.string() }),
  outputSchema: z.object({ displayed: z.boolean() }),
})
```

```typescript
// tools/server.ts
import { getProductsDef } from './definitions'

export const getProducts = getProductsDef.server(async ({ query, limit }) => {
  const results = await db.products.search(query, { limit: limit ?? 10 })
  return {
    products: results.map((p) => ({ id: p.id, name: p.name, price: p.price })),
  }
})
```

```typescript
// api/chat/route.ts
import { chat, toServerSentEventsResponse } from '@tanstack/ai'
import { openaiText } from '@tanstack/ai-openai'
import { getProducts } from '@/tools/server'
import { updateCartUIDef } from '@/tools/definitions'

export async function POST(request: Request) {
  const { messages } = await request.json()
  const stream = chat({
    adapter: openaiText('gpt-5.5'),
    messages,
    tools: [getProducts, updateCartUIDef], // server tool + client definition
  })
  return toServerSentEventsResponse(stream)
}
```

```typescript
// app/chat.tsx
import {
  useChat,
  fetchServerSentEvents,
  clientTools,
  createChatClientOptions,
  type InferChatMessages,
} from "@tanstack/ai-react";
import { updateCartUIDef } from "@/tools/definitions";
import { useState } from "react";

function ChatPage() {
  const [cartCount, setCartCount] = useState(0);

  const updateCartUI = updateCartUIDef.client((input) => {
    setCartCount(input.itemCount);
    return { displayed: true };
  });

  const tools = clientTools(updateCartUI);
  const chatOptions = createChatClientOptions({
    connection: fetchServerSentEvents("/api/chat"),
    tools,
  });
  const { messages, sendMessage } = useChat(chatOptions);
  // InferChatMessages ties part types to the configured tools when needed:
  // type Messages = InferChatMessages<typeof chatOptions>

  return (
    <div>
      <span>Cart: {cartCount}</span>
      {messages.map((msg) => (
        <div key={msg.id}>
          {msg.parts.map((part) => {
            if (part.type === "text") return <p>{part.content}</p>;
            if (part.type === "tool-call") {
              return <div key={part.id}>Tool: {part.name} ({part.state})</div>;
            }
            return null;
          })}
        </div>
      ))}
    </div>
  );
}
```

## Core Patterns

### Generic middleware interrupts

Use `defineInterrupt()` when middleware needs typed data from the client. This
does not replace `needsApproval`. Tool approval asks whether a tool can run.
Generic interrupts ask for application data at a chat lifecycle boundary.

Define the interrupt once. Register it with both `chat({ interrupts })` and
`useChat({ interrupts })`. Emit it only from `onInterruptBoundary`, then read
the typed result in `onInterruptResolution`.

```typescript
import { defineInterrupt, type ChatMiddleware } from '@tanstack/ai'
import { z } from 'zod'

const reviewPlan = defineInterrupt({
  id: 'review-plan',
  payloadSchema: z.object({ title: z.string() }),
  responseSchema: z.object({ approved: z.boolean() }),
})

const reviewMiddleware: ChatMiddleware<unknown, typeof reviewPlan> = {
  onInterruptBoundary(ctx) {
    if (ctx.phase !== 'beforeTools') return
    return {
      interrupts: [
        reviewPlan.interrupt({
          key: 'release-plan',
          reason: 'review-required',
          message: 'Approve this plan?',
          payload: { title: 'Release plan' },
        }),
      ],
    }
  },
  onInterruptResolution(_ctx, resumedInterrupts) {
    for (const result of resumedInterrupts.for(reviewPlan)) {
      if (result.status === 'resolved' && !result.response.approved) {
        return { toolResume: 'stop' }
      }
    }
  },
}
```

Several middleware can request generic interrupts at one boundary. They share
one AG-UI interrupt batch with tool approvals. A continuation starts only after
the client resolves or cancels every bound item. `stop` is more restrictive than
`cancel`, which is more restrictive than `continue`.

Do not emit raw AG-UI interrupt events from middleware. Use the boundary hook
so the engine creates one terminal event and persistence records the batch.

### Pattern 1: Server-Only Tool

Define with `toolDefinition()`, implement with `.server()`, pass to `chat({ tools })`.
The server executes it automatically. The client never runs code for this tool.

```typescript
import { toolDefinition } from '@tanstack/ai'
import { z } from 'zod'

const getUserDataDef = toolDefinition({
  name: 'get_user_data',
  description: 'Look up user by ID',
  inputSchema: z.object({
    userId: z.string().meta({ description: "The user's ID" }),
  }),
  outputSchema: z.object({ name: z.string(), email: z.string() }),
})

const getUserData = getUserDataDef.server(async ({ userId }) => {
  const user = await db.users.findUnique({ where: { id: userId } })
  return { name: user.name, email: user.email }
})

// In your route handler:
const stream = chat({
  adapter: openaiText('gpt-5.5'),
  messages,
  tools: [getUserData],
})
```

### Pattern 2: Client-Only Tool

Pass the bare definition (no `.server()`) to `chat({ tools })` so the LLM knows
about it. Pass