> ## 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.

# LiteLLM

> completion() and acompletion() with the context, and a CustomLogger that records the turns, for any provider routed through LiteLLM.

`niadra.integrations.litellm` brings `completion()` and `acompletion()`, which are `litellm.completion` and `litellm.acompletion` with the context, and `NiadraLogger`, a LiteLLM `CustomLogger` that records the answer of every call made inside a conversation. Python only; it covers any provider LiteLLM routes.

## Install

```sh theme={null}
pip install 'niadra[litellm]'   # litellm 1.102 or newer, below 2
```

## The five primitives

| Primitive    | How the adapter wires it                                                                                                                                                                                                                                |
| ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Context      | Inside a conversation or task block, `completion()` and `acompletion()` put the pack (after the agent's notes, with `agent_memory=`) after your leading system messages and the `turn_block` at the end. Every other argument goes to LiteLLM unchanged |
| Turns        | The answer (streamed or not) is recorded as the agent's turn with the usage LiteLLM normalized. `NiadraLogger`, in `litellm.callbacks`, does the same for code that calls LiteLLM (or its `Router`) directly; it never injects                          |
| Tools        | `conversation.tools()` hands over the kit for `tools=`                                                                                                                                                                                                  |
| Verification | `conversation.verify()` before the call                                                                                                                                                                                                                 |
| Handoff      | `conversation.handoff()` where your flow transfers                                                                                                                                                                                                      |

## Minimal example

```python theme={null}
"""Any provider through LiteLLM with the customer's context."""

from niadra import Niadra, phone
from niadra.integrations.litellm import completion

niadra = Niadra(channel="chat")

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

The same code is in `examples/litellm_completion.py`.

## Limits

* Outside a block, `completion()` and `acompletion()` are LiteLLM's own calls.
* `NiadraLogger` only records; for the pack to go in, use `completion()`.
* The `litellm` extra pins versions incompatible with `openai-agents`; install one per environment.
* Tested against `litellm` 1.102 with the transport replaced and Niadra on the emulator.

## Next steps

<CardGroup cols={2}>
  <Card title="OpenAI and Azure OpenAI" href="/en/integrations/openai">
    the client `wrap()`, with no router in between.
  </Card>

  <Card title="Python SDK" href="/en/sdk/python">
    the conversation, `mark_injected()` and the provider's usage.
  </Card>
</CardGroup>