Platform · Anthropic

Observe AI agents that call Anthropic.

Your agent calls Claude Opus, Sonnet, or Haiku via the Messages API. Trefur auto-instruments the SDK and captures each call as a span on the agent trace — with prompts, completions, tool cycles, and any retries in between.

What you see.

Model and version per Messages call — Claude Opus, Sonnet, Haiku.

Input and output tokens, USD attributed per Anthropic pricing.

Tool use cycles — tool requested, tool input, tool result, follow-up completion.

Stream vs non-stream captured as one completed span with the full text.

Stop reason — end_turn, max_tokens, stop_sequence, tool_use.

Rate limits (429), overloaded errors, context-length warnings.

Multi-turn message threads carried as the prompt content on the span.

Why it matters.

For engineers

One line at startup auto-instruments the Anthropic SDK. Existing Messages and tool-use code keeps working; the trace shows every call with the prompt the model saw and the completion it returned.

For platform teams

Roll up Anthropic usage per agent, per environment. One trace shape across Anthropic, OpenAI, Bedrock, Azure so dashboards keep working when teams swap vendors.

For business operators

Know which workflows depend on Claude models and what they cost. Catch silent model upgrades or downgrades that drift cost-per-run.

For SRE

Anthropic overloaded and rate-limit errors land on the span. Envelope alerts catch retry-rate spikes and route to oncall with the trace already attached.

How it works.

The Trefur SDK wraps the Anthropic client. TrefurObserve.init() and patch_anthropic(obs) at startup register the hooks; every messages.create call emits a span with the prompt, the completion, tokens, USD per current pricing, and any tool-use cycle as children. The collector can run inside your network if you want prompt content to stay local.

See what your Claude agents are actually doing.

Free tier. No card required. First trace in under five minutes.

Start free See all platforms