← Back to Discover

traceAI

Open-source libraries that record LLM calls, prompts, token counts, retrieval steps and agent decisions as OpenTelemetry traces, with packages for Python, TypeScript, Java and C# applications.

Observability & evals
Visit website ↗

What it does

traceAI adds tracing to applications built on language models. It captures each LLM call, prompt, token count, retrieval step, tool call and agent decision as structured OpenTelemetry data, so you can see where an answer went wrong. Instrumentation is added with a few lines of code next to your existing client. Packages exist for Python, TypeScript, Java and C#, and the README lists integrations for model providers, agent frameworks, vector databases and other libraries. Because the traces follow OpenTelemetry conventions, they can be sent to any compatible backend, such as Datadog, Grafana, Jaeger or Future AGI, instead of requiring a new dashboard.

Capabilities

  • Drop-in instrumentation for 50+ AI frameworks and libraries
  • Records prompts, completions, token counts, model parameters and tool calls
  • Works with any OpenTelemetry-compatible backend
  • Shared naming conventions for spans covering models, agents, tools, retrieval and vector stores
  • Packages for Python, TypeScript, Java and C# with consistent APIs
  • Supports async and streaming calls

When to use it

  • You need to debug a failing LLM or agent pipeline by inspecting each step
  • Your team already uses an OpenTelemetry backend and does not want another vendor dashboard
  • You run AI features in Java or C# as well as Python or TypeScript