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

# Anthropic

> Using Cognisafe with Anthropic — direct mode.

## How it works

Anthropic uses **direct mode**. Because Anthropic's API is not OpenAI-compatible, Cognisafe wraps the `messages.create` method on the Anthropic client rather than routing through the proxy. After the model responds:

1. Your code receives the response immediately
2. The SDK fires a background task that ships the request and response payload to `POST /internal/log` on the FastAPI backend
3. The backend queues a scoring job in Redis

No proxy is involved. The background shipping adds no perceptible latency to your application.

## Installation

```bash theme={null}
pip install "cognisafe[anthropic]"
```

## Setup

```python theme={null}
import cognisafe
import anthropic

cognisafe.configure(
    api_key="csk_your_key_here",
    project_id="my-app",
)
cognisafe.patch_anthropic()

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "Explain quantum entanglement in one paragraph."}
    ],
)

print(message.content[0].text)
```

## Supported capabilities

| Capability                       | Supported | Notes                                   |
| -------------------------------- | --------- | --------------------------------------- |
| Messages API (`messages.create`) | Yes       | Full request and response logged        |
| Streaming                        | Yes       | Captured after stream completes         |
| System prompts                   | Yes       | Included in logged payload              |
| Tool use                         | Yes       | Tool definitions and results logged     |
| Vision (image inputs)            | Yes       | Image payloads included in request body |

## Scoring note

Safety scoring always uses the model configured via `SCORER_MODEL` (default: `gpt-4o-mini`), regardless of which Claude model was used for the original call. PyRIT's scorers are model-agnostic — they receive the text of the prompt and response, not a reference to the originating model.

<Note>
  Claude models and the scoring model are completely independent. You can use `claude-opus-4-5` for your product and GPT-4o mini as the safety scorer with no conflict.
</Note>

## Async client

`patch_anthropic()` also wraps the async client:

```python theme={null}
import asyncio
import cognisafe
import anthropic

cognisafe.configure(api_key="csk_...", project_id="my-app")
cognisafe.patch_anthropic()

async def main():
    client = anthropic.AsyncAnthropic()
    message = await client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=512,
        messages=[{"role": "user", "content": "Hello."}],
    )
    print(message.content[0].text)

asyncio.run(main())
```