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

# Anthropic

> The wrap() of the Anthropic client: the pack in system, the turn_block on the last user message and the usage with cache reads and writes.

`wrap()` (Python) and `wrapAnthropic()` (TypeScript) wrap the Anthropic client, including the Bedrock and Vertex clients of the same SDK in Python, so `messages.create` and `messages.stream` get the context and record the answer.

## Install

<CodeGroup>
  ```sh Python theme={null}
  pip install 'niadra[anthropic]'   # anthropic 1.8 or newer, below 2
  ```

  ```sh TypeScript theme={null}
  npm install @niadra/sdk @anthropic-ai/sdk   # @niadra/sdk/anthropic
  ```
</CodeGroup>

The TypeScript integration ships with `@niadra/sdk` 0.3.0, ready on `main` and on npm when it is published; until then the npm package is 0.1.1.

## The five primitives

| Primitive    | How the adapter wires it                                                                                                                                                                                                                                                                                                                                                                               |
| ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| Context      | The pack (after the agent's notes, with `agent_memory=`) goes in `system`, as a text block after your own system text; it carries `cache_control` only when your system blocks already do, so it joins a prefix you chose to cache. The `turn_block` goes as a text block at the end of the last user message, tagged as data, or as a user message of its own when the last one is not the customer's |
| Turns        | The answer (its text blocks) is recorded as the agent's turn with Anthropic's usage: `input_tokens`, `cache_read_input_tokens` and `cache_creation_input_tokens`. In TypeScript, the newest user message is also recorded as the customer's turn                                                                                                                                                       |
| Tools        | `kit.anthropic_definitions()` (Python) gives the kit in the `name`, `description`, `input_schema` shape; the wrapper does not touch the tools                                                                                                                                                                                                                                                          |
| Verification | `conversation.verify()` before the call                                                                                                                                                                                                                                                                                                                                                                |
| Handoff      | `conversation.handoff()` where your flow transfers                                                                                                                                                                                                                                                                                                                                                     |

## Minimal example

<CodeGroup>
  ```python Python theme={null}
  """Anthropic's Messages API with the customer's context in the system prompt."""

  from anthropic import Anthropic

  from niadra import Niadra, phone
  from niadra.integrations.anthropic import wrap

  niadra = Niadra(channel="chat")
  claude = wrap(Anthropic())

  with niadra.conversation("thread-81", subject=phone("+5511912345678")) as conversation:
      conversation.customer("Where is my replacement lid?")
      message = claude.messages.create(
          model="claude-sonnet-4-5",
          max_tokens=512,
          system=[{"type": "text", "text": "You are Acme's agent.", "cache_control": {"type": "ephemeral"}}],
          messages=[{"role": "user", "content": "Where is my replacement lid?"}],
      )
      print(message.content[0].text)
  ```

  ```typescript TypeScript theme={null}
  import Anthropic from "@anthropic-ai/sdk";
  import { Niadra, handles } from "@niadra/sdk";
  import { wrapAnthropic } from "@niadra/sdk/anthropic";

  const niadra = new Niadra();
  const anthropic = new Anthropic();

  /** One customer message in; `userId` comes from your session, never from the model. */
  export async function reply(userId: string, chatId: string, text: string): Promise<string> {
    const convo = niadra.conversation({ subject: handles.appUserId(userId), channel: "web_chat", conversation_id: chatId });
    const claude = wrapAnthropic(anthropic, convo, { verify: { method: "login", level: "V2" } });
    const message = await claude.messages.create({
      model: "claude-sonnet-4-5",
      max_tokens: 1024,
      system: "You are Acme's support agent. Be brief.",
      messages: [{ role: "user", content: text }],
    });
    return message.content.map((block) => (block.type === "text" ? block.text : "")).join("");
  }
  ```
</CodeGroup>

The same code is in `examples/anthropic_messages.py` and `examples/anthropic.ts`.

## Limits

* Outside a conversation or task block, calls pass through untouched, and nothing the wrapper does can fail the call.
* In TypeScript, for `messages.stream()` prepare the body with `anthropicParams()` and record the final message with `recordAnthropic()`.
* The wrapper adds no `cache_control` where you use no prompt caching: it never changes your caching strategy.
* Tested against `anthropic` 1.8 and `@anthropic-ai/sdk` 0.128 with the transport replaced and Niadra on the emulator.

## Next steps

<CardGroup cols={2}>
  <Card title="Amazon Bedrock" href="/en/integrations/bedrock">
    the same for the Converse API.
  </Card>

  <Card title="Context and views" href="/en/concepts/context">
    the target model and the cache floors.
  </Card>
</CardGroup>