Open-Source Tracy Brings AI Observability to Kotlin

As companies move large language models and AI agents into production, debugging has grown harder because model requests, tool calls and custom business logic often span multiple systems. Tracy, an open-source Kotlin library built on OpenTelemetry, addresses that gap without tying developers to a single AI framework. It is designed to expose execution timing and causal links across AI components and agent states, helping teams diagnose failures, assess performance and support production evaluations.

Tracy was unveiled in March 2026, with iThome reporting the launch on March 13. The library offers three high-level tracing methods: scoped blocks using withSpan, instrumentation for LLM clients, and @Trace annotations for functions or interface methods. It works with OkHttp and Ktor, as well as OpenAI, Anthropic and Gemini clients, and exports data to OpenTelemetry-compatible backends such as Jaeger, Zipkin and Grafana or specialist platforms including Langfuse and W&B Weave. By default, Tracy captures metadata such as the API, model and parameters; logging prompts and outputs requires explicit activation.