Developers are rapidly shifting toward agent-assisted coding tools like Cursor, Claude Code and ChatGPT to move faster from conception to production. But one part of the process refuses to modernize: debugging. Production observability still lives in dashboards and terminals outside the coding environment, forcing developers to jump between multiple tools to diagnose issues. The friction remains unchanged, even as the tooling evolves.
Traceloop is aiming to close that gap with the recently released open-source MCP (Model Context Protocol) server for OpenTelemetry that allows developers to query and analyze production traces directly inside MCP-enabled environments. It brings observability into the same environment where code is written, reviewed and executed.
The development comes after two years spent supporting teams debugging LLM-powered apps and distributed systems in production. The pain points were consistent: every debugging session required bouncing between Grafana, Jaeger, Tempo, Datadog and internal tools to understand failures and latency. Observability was siloed, even as the rest of the development flow shifted to integrated tooling.
Traceloop’s new server changes the model. Unlike Datadog’s proprietary MCP implementation, the server is open-source and designed for multi-backend compatibility. It supports Jaeger, Grafana Tempo, Traceloop, and other OpenTelemetry-based systems, and works across the major MCP-enabled clients, including Claude Desktop, Cursor, Codeium, Gemini CLI and ChatGPT.
The tool is also LLM-aware. It incorporates OpenLLMetry semantic conventions so developers can analyze token usage, compare model performance, inspect expensive inference calls and detect latency issues—all without leaving their IDE or assistant. Debugging AI applications becomes a first-class workflow, not an afterthought.
“We kept seeing the same problem across every team we worked with,” said Nir Gazit, co-founder of Traceloop. “Developers write code in one place, then investigate failures in another. The context switching is where debugging dies. MCP gives us a way to bring observability into the developer’s world instead of forcing them to jump into vendor dashboards. Our goal is to make production analysis feel as natural as running a unit test.”
The server offers tools for common workflows: finding slow or expensive traces, identifying errors, comparing model efficiency, tracking token consumption, and debugging distributed systems. It supports both stdio and HTTP/SSE transports, enabling local debugging or remote deployments. Installation takes a single command using pipx or uv, with no global dependencies.
Traceloop is positioning the project as part of a broader shift. Modern systems blend microservices, multiple inference models, multi-cloud deployments and OpenTelemetry observability standards. Debugging those environments through fragmented dashboards slows teams down.
