Appearance

Anthropic ​

examples/03_anthropic.py demonstrates the Anthropic integration. As with the OpenAI example, no agentsight call appears near the provider: Messages.create and the client.messages.stream(...) manager are both wrapped, so a streamed answer produces one span with real token counts rather than nothing.

anthropic is not a dependency of this project, so the script reports and exits when the package is missing rather than faking it:

bash
pip install anthropic
python examples/03_anthropic.py
AGENTSIGHT_EXAMPLES_OFFLINE=1 python examples/03_anthropic.py

What it runs ​

  • A plain message — client.messages.create() between the user and agent messages. The stub's usage block carries cache_creation_input_tokens and cache_read_input_tokens alongside plain input: Anthropic prices cache writes above plain input and cache reads far below it, so neither can be folded into input_tokens — the span keeps all four counts separate. See Tokens & Cost.
  • A streamed message — through the with client.messages.stream(...) manager. The manager's __enter__ is patched, not just create(): a stream consumed through with reports usage only at the end, and closing the block is what completes the span.
  • A failure — a request for a model that does not exist; the span records the error.

Two details in the script are worth stealing for real integrations. It reads the answer with a helper that takes the first text block, not content[0] — on a thinking-capable model the first block is a ThinkingBlock, and the reasoning is not what the customer was told. And it sets max_tokens=1024 rather than something small: on a thinking-capable model the reasoning block spends from the same budget as the answer, and a low cap gets eaten whole by thinking — the span still records the capped output tokens, but the reply comes back empty.

What to look for in the output ​

Three llm spans in examples/traces/03_anthropic/. The plain message's span carries all four token counts; the streamed one is marked streamed with usage assembled from the stream's start and end events.