LLM Monitoring Overview

LLM Monitoring gives you distributed tracing for AI workloads: the model calls that happen, the agents and tools they invoke, the tokens they consume, and what they cost. This page explains what Atatus collects and helps you pick the right way to send it. Traces appear on the LLM page in Atatus, separate from your APM services. Each trace shows the full call tree, so you can see a request or a prompt at the top, the model calls beneath it, and the tools each model call triggered, with token counts and latency at every level.

Two ways to send LLM data

Atatus collects LLM traces from two kinds of producer. They work side by side, so you can use both in the same account and both feed the same LLM pages.

Producer	What it monitors	How you install it
Coding agents	The AI coding assistants your team uses, such as Claude Code, Codex, and Cursor. Each prompt becomes a trace.	Install the tracing package on the developer’s machine. See Coding agents.
APM agents	LLM calls made by your own application, such as an API, a background worker, or a RAG pipeline. Traces link to the surrounding request.	Turn on an option in an Atatus APM agent you already run.

This section documents the coding agent path. If you want visibility into how your developers use AI coding tools, start with Coding agents.

Instrumenting your own application is supported in the Atatus Python agent. See the Python agent overview to get started.

What a trace contains

Whichever producer sends it, a trace is a tree of spans, and every span has a kind:

Span kind	Meaning
`chat`, `completion`	A single model call. Carries the model name, provider, token counts, and cost.
`agent`	An autonomous unit that decides its own next step.
`tool`	One tool or function invocation.
`chain`	A sequence of steps grouped together.
`workflow`, `task`	Groupings you define in your own code with decorators.

For the full reference, see Span kinds. For how spend is calculated, see Cost and tokens.

Where your data goes

Both producers send traces to https://otel-rx.atatus.com over OTLP/HTTP, authenticated with your Atatus license key. There is no separate collector to run.

On-premise installations use their own endpoint instead of the hosted one. Check with your Atatus administrator for the correct address.

How projects appear in the UI

Projects are typed, and the type decides where a project shows up:

A coding-agent project appears on the LLM page only. It never appears under APM, because it has no web transactions to show there.
An APM service with LLM monitoring enabled appears in both places, because it has ordinary request traces as well as LLM spans.

If you installed a coding agent and cannot find it, look on the LLM page rather than APM. This is the most common reason people think the install did not work.

Next steps

Pick the path that matches what you want to monitor:

Getting started covers what you need before you install.
Coding agents overview covers the ten supported tools.
Span kinds explains how to read a trace.
Evaluators score your traces with an LLM-as-a-judge, once traces are arriving.

​Two ways to send LLM data

​What a trace contains

​Where your data goes

​How projects appear in the UI

​Next steps

Two ways to send LLM data

What a trace contains

Where your data goes

How projects appear in the UI

Next steps