Claude Code is a black box. Here is how to trace its tool calls and LLM requests (www.arthur.ai)

A new open-source integration has been unveiled that enhances the visibility of Claude Code, a powerful coding tool that autonomously executes LLM requests and tool invocations. This integration sends structured OpenInference traces to Arthur Engine, allowing users to monitor every step in their Claude Code sessions, from user prompts to tool calls and errors. This capability addresses the "black box" nature of Claude Code, providing crucial insights into its operations, including which files were accessed, the success of tool invocations, and any sensitive information that may leak into prompts. The integration is significant for the AI/ML community, particularly for developers and engineering teams using Claude Code in production. By effectively logging and tracing actions within Claude Code, teams can mitigate risks involving sensitive data exposure, track token usage and costs, and ensure compliance with usage guidelines. The integration not only brings greater observability to coding practices but also supports continuous evaluation and testing of code. With its straightforward installation process, users can easily implement this tool to enhance their workflows and understanding of Claude Code's operations, making it a valuable addition to any engineering toolkit.