# LiteLLM with Zurelay

> One OpenAI-style API for every provider: a Python SDK and a proxy server. Made by BerriAI.

Source: https://zurelay.com/docs/integrations/litellm

The examples use `claude-opus-5-5`. Any model ID from [the models page](https://zurelay.com/docs/models) works.

## Set it up

1. Install it: `pip install litellm` for the Python SDK, or `pip install 'litellm[proxy]'` for the proxy server.
2. Prefix Zurelay model IDs with `openai/` and set `api_base` to the base URL, with `/v1`.
3. For the proxy, list the models in `config.yaml` and run `litellm --config config.yaml`. It serves `http://localhost:4000` to OpenAI and Anthropic clients alike.
4. Keep `use_chat_completions_api` and `use_chat_completions_url_for_anthropic_messages` on. Without them LiteLLM turns Responses and Anthropic Messages requests (Claude Code’s) into calls to OpenAI’s Responses API, which Zurelay doesn’t serve.

## Configuration

**config.yaml** (`config.yaml`):

```yaml
model_list:
  - model_name: claude-opus-5-5
    litellm_params:
      model: openai/claude-opus-5-5
      api_base: https://api.zurelay.com/v1
      api_key: os.environ/ZURELAY_API_KEY
      use_chat_completions_api: true
    model_info:
      input_cost_per_token: 0.0000014
      output_cost_per_token: 0.000007
      cache_read_input_token_cost: 0.00000007
  # Every other Zurelay model, as zurelay/<model ID>
  - model_name: zurelay/*
    litellm_params:
      model: openai/*
      api_base: https://api.zurelay.com/v1
      api_key: os.environ/ZURELAY_API_KEY
      use_chat_completions_api: true

litellm_settings:
  use_chat_completions_url_for_anthropic_messages: true
  check_provider_endpoint: true

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY
```

**Proxy** (`terminal`):

```bash
pip install 'litellm[proxy]'
export LITELLM_MASTER_KEY="sk-choose-a-long-random-key"
litellm --config config.yaml

# from another terminal
curl http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "claude-opus-5-5", "messages": [{"role": "user", "content": "Hello"}]}'
```

**Python SDK** (`app.py`):

```python
import os
from litellm import completion

response = completion(
    model="openai/claude-opus-5-5",
    api_base="https://api.zurelay.com/v1",
    api_key=os.environ["ZURELAY_API_KEY"],
    messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)
```

> **Note:** LiteLLM prices requests from its own list: at the lab’s list price for models it knows, and not at all for the rest. `model_info` puts Zurelay’s prices in its spend tracking. For Claude through LiteLLM’s `anthropic/` provider instead, use `https://api.zurelay.com` with no `/v1`; LiteLLM adds `/v1/messages` itself.

We ran this setup end to end with DeepSeek and Claude. The [LiteLLM guide](https://zurelay.com/guides/litellm) has every error we hit, how to fix it, and what each test task cost.

[LiteLLM’s own docs](https://docs.litellm.ai/docs/providers/openai_compatible)