Lunary Python SDK

📈 Python monitoring for AI apps and agent

website - docs -

---

Use it with any LLM model and custom agents (not limited to OpenAI).

To get started, get a project ID by registering here.

🛠️ Installation

pip install lunary

🤖 Anthropic Integration

from anthropic import Anthropic
import lunary

client = Anthropic()
lunary.monitor(client)

parsed = client.messages.parse(
    model="claude-sonnet-4-5-20250929",
    max_tokens=256,
    messages=[{"role": "user", "content": "Return JSON with an answer field for 2 + 2."}],
    output_format=dict,
)

with client.messages.stream(
    model="claude-sonnet-4-5-20250929",
    max_tokens=256,
    messages=[{"role": "user", "content": "Stream one sentence about Lunary."}],
) as stream:
    for event in stream:
        pass

Diagnosing event delivery errors

Keep LUNARY_VERBOSE unset or false in production and capture ERROR logs from lunary.consumer. Each failed send produces one line explaining what happened and whether the SDK will retry. It includes the exception type, HTTP status when available, and the number of events in the batch.

Error sending events. Invalid API key. Batch will not be retried. error_type=HTTPError http_status=401 batch_size=3.
Error sending events. The endpoint rate-limited the request. Batch will not be retried. error_type=HTTPError http_status=429 batch_size=3.
Error sending events. The ingestion service or gateway is unavailable. Batch will not be retried. error_type=HTTPError http_status=503 batch_size=3.
Error sending events. The endpoint did not respond in time. Batch will not be retried. error_type=ReadTimeout http_status=n/a batch_size=3.

For recognized API key, project and billing failures, the SDK uses the API's fixed explanation. For other responses it describes the HTTP status or exception type. Normal logs omit arbitrary server text, request URLs, credentials, event payloads and tracebacks. Verbose mode retains full tracebacks and can log event contents. Share the timestamp, SDK version and diagnostic line with support.

The consumer removes a batch from its in-memory queue before sending and does not automatically retry a failed send. An error therefore does not mean the batch will eventually be delivered; after a connection failure, server acceptance may also be uncertain. This diagnostic change does not add retries, a timeout, persistence, or recovery of missing traces.

📖 Documentation

Full docs are available here.