Wrappers provide the simplest way to add tracing to your LLM calls. Wrap your provider client once and every API call is automatically traced — no other code changes required.
The wrapper returns the same client type with tracing injected. Your existing code continues to work identically.
All wrappers automatically capture: input messages, output messages, model name, token usage (prompt, completion, reasoning, cache tokens where available), model parameters, streaming support, time-to-first-token for streaming, and error details.
Comparison
| Wrapper | Install Extra | Provider Methods Traced |
|---|
wrap_openai | pandaprobe[openai] | chat.completions.create, responses.create |
wrap_anthropic | pandaprobe[anthropic] | messages.create, messages.stream |
wrap_gemini | pandaprobe[gemini] | models.generate_content, models.generate_content_stream |
Install only the provider extras you use so dependency resolution stays minimal in production images and lockfiles.
Quick example
from pandaprobe.wrappers import wrap_openai
from openai import OpenAI
client = wrap_openai(OpenAI())
# Use client exactly as before — all calls are traced automatically
Wrappers work seamlessly with manual instrumentation. If a wrapper call happens inside a pandaprobe.start_trace() or @pandaprobe.trace context, the LLM span is automatically nested as a child span.
Provider guides
OpenAI
Chat Completions and Responses API, streaming, and tool spans
Anthropic
Messages API, streaming patterns, and extended thinking
Google Gemini
generate_content, async, streaming, and thinking mode
