Watch the agent. Not the spinner.

What is it?

A local OpenTelemetry trace collector and viewer for AI-agent execution. A trace is a record of the steps a program took.

What problem does it solve?

A local view into the calls, prompts and pauses behind an AI run.

What does it replace?

Potentially a heavier tracing service during local development, when all you need is to understand one run.

What doesn’t it replace?

Production monitoring, alerting, or instrumentation inside the application being observed.

Who is it for?

People building agents who need to see what happened between request and response.

The review notebook

A spinner tells you that something is happening. It does not tell you whether the model is thinking, a tool is waiting or a request has quietly failed. Local tracing promises a much more useful conversation with the program.

The small setup is the attraction here. We would start with a familiar workload and compare its known steps with what appears in the viewer. Missing steps matter just as much as a pleasing timeline. Cost estimates also deserve to be read as estimates, not billing records.

Known catches & open questions

The default trace buffer is in memory. Persistence is optional, and useful visibility still depends on what your application emits.

Current judgement

This is a promising lead, not a recommendation. We will only make a hands-on judgement after the stated test has been run.

Next on the bench

Send a known synthetic trace, restart the collector, and compare default retention with explicitly enabled persistence.

Sources & evidence

Project repository and README ↗
Release history ↗

Public material checked 5 September 2026. Version labels are the latest stable release returned at that check, where available. Platform descriptions are upstream claims, not Toolglass acceptance tests.