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

# Claude Code / Anthropic SDK

> Use @sigilcore/agent-hooks with Claude Code and the Anthropic SDK to gate tool calls before execution.

## Installation

```bash theme={null}
npm install @sigilcore/agent-hooks
```

## Adapter Status

Claude Code and the Anthropic SDK already have a dedicated package export:
`checkAnthropicToolUse`. Use the generic model-budget helpers
`recordModelUsage` and `checkModelBudget` around Anthropic SDK responses when
the host owns the model loop and can read response usage.

HTTP note: emit a typed `http` intent only when the intercepted tool input explicitly contains a valid method. Otherwise retain `web_fetch`; never infer `GET`.

## Usage

```typescript theme={null}
import { checkAnthropicToolUse } from '@sigilcore/agent-hooks';

const config = {
  apiKey: process.env.SIGIL_API_KEY!,
  agentId: 'my-claude-agent',
};

// In your PreToolUse hook:
const rejection = await checkAnthropicToolUse(toolUseBlock, config);
if (rejection) {
  // Feed rejection back to Claude as a tool_result error
  return rejection;
}
// Otherwise, let the tool execute normally
```

## Tool Name Mapping

`checkAnthropicToolUse` maps Anthropic tool names to Sigil action types automatically:

| Anthropic Tool          | Sigil Action         |
| ----------------------- | -------------------- |
| `Bash`, `bash`          | `bash`               |
| `WebSearch`, `WebFetch` | `web_fetch`          |
| `Write`, `Edit`         | `file_write`         |
| `computer`              | `bash`               |
| Any other tool          | lowercased tool name |

## What Gets Sent to Sigil

For each tool call, the following intent is submitted to `/v1/authorize`:

```typescript theme={null}
{
  action: 'bash',                    // mapped from tool name
  command: block.input['command'],   // bash only
  url: block.input['url'],           // web_fetch only
  path: block.input['path'],         // file_write only
  metadata: block.input,             // full input for custom rules
}
```

## Rejection Response

When Sigil denies or holds an action, `checkAnthropicToolUse` returns a `tool_result` error block that Claude understands:

```typescript theme={null}
{
  type: 'tool_result',
  tool_use_id: block.id,
  content: JSON.stringify({
    sigil_decision: 'DENIED',
    sigil_error_code: 'SIGIL_POLICY_VIOLATION_BLOCKED_COMMAND',
    sigil_message: 'Command contains blocked string: rm -rf',
    sigil_action_taken: 'halted',
    sigil_next_steps: 'Do not attempt to reframe or retry this action.',
  }),
  is_error: true,
}
```

Claude will receive this as a tool error and adjust its behavior accordingly.

## Configuration

| Field       | Type       | Required | Default                      | Description                    |
| ----------- | ---------- | -------- | ---------------------------- | ------------------------------ |
| `apiKey`    | `string`   | Yes      | —                            | Sigil API key (`sk_sigil_...`) |
| `apiUrl`    | `string`   | No       | `https://sign.sigilcore.com` | Sigil Sign endpoint            |
| `agentId`   | `string`   | No       | `'agent'`                    | Agent identifier               |
| `onDenied`  | `function` | No       | —                            | Called when action is denied   |
| `onPending` | `function` | No       | —                            | Called when action is held     |
| `onError`   | `function` | No       | —                            | Called on network error        |