> ## Documentation Index
> Fetch the complete documentation index at: https://docs.compose.market/llms.txt
> Use this file to discover all available pages before exploring further.

# Tools

> How agents use tools — models, connectors, agents, search, and catalog.

When an agent runs a task, it can call tools to extend its capabilities. During a stream, every tool call produces a `tool_start` event followed by a `tool_end` event. You can parse these into typed objects to build UIs, logs, or audit trails.

## Tool types

Agents on Compose have access to five tool families:

| Tool       | Tool name         | What it does                                                       |
| ---------- | ----------------- | ------------------------------------------------------------------ |
| Models     | `models_call`     | Calls a model for text, image, audio, or video generation          |
| Connectors | `connectors_find` | Discovers external integrations like MCP servers and onchain tools |
| Agents     | `agent_find`      | Finds other deployed agents for delegation                         |
| Search     | `search_call`     | Runs a web search through a chosen provider                        |
| Catalog    | `query_catalog`   | Queries the catalog when the right tool family is unclear          |

A sixth tool name, `swarm_delegate`, belongs to the Agents family. It delegates a task to another agent after discovery. See [Agents](/sdk/tools/agents).

## Parsing tool events

Use `parseToolEvent()` to turn a raw `tool_start` or `tool_end` event into a typed object. Each tool family has its own shape, so you can inspect the parsed result to know which tool ran.

```typescript theme={null}
import { parseToolEvent } from "@compose-market/sdk";

for await (const event of stream) {
  if (event.type === "tool_start" || event.type === "tool_end") {
    const tool = parseToolEvent(event);
    if (!tool) continue;

    if (tool.start) {
      console.log(`Tool started: ${event.toolName}`);
    } else if (tool.error) {
      console.error(`Tool failed: ${tool.error}`);
    } else {
      console.log(`Tool finished: ${event.toolName}`);
    }
  }
}
```

## Common shape

Every tool event shares the same core fields:

| Field    | Type             | Meaning                                       |
| -------- | ---------------- | --------------------------------------------- |
| `start`  | `boolean`        | `true` on `tool_start`, `false` on `tool_end` |
| `action` | `string`         | What the tool was asked to do                 |
| `result` | `object \| null` | The result payload (only on `tool_end`)       |
| `error`  | `string \| null` | Error message if the tool failed              |

Each family adds one more field that identifies the subject — `model`, `connector`, `agent`, `mode`, or `domain`. The pages below cover each one in detail.

## Next

<CardGroup>
  <Card title="Models" icon="sparkles" href="/sdk/tools/models">Model discovery and invocation.</Card>
  <Card title="Connectors" icon="plug" href="/sdk/tools/connectors">Connector discovery — MCP and onchain.</Card>
  <Card title="Agents" icon="bot" href="/sdk/tools/agents">Agent discovery and delegation.</Card>
  <Card title="Search" icon="search" href="/sdk/tools/search">Web search through the stream.</Card>
  <Card title="Catalog" icon="book-open" href="/sdk/tools/catalog">Catalog queries for tool selection.</Card>
</CardGroup>