This page takes you from nothing to your first LLM trace. It covers tracing the AI coding assistants your team uses, which is the path most people start with.

​
What you need

Two things, both quick to get:
  • Your Atatus license key, from Settings then License key in the Atatus dashboard.
  • A project name that you choose. It groups the data and names the project in the UI.
You also need Python 3.9 or later on the machine you are tracing, and one of the ten supported coding agents installed.
The project name is the grouping key, so it decides how your data is organised. Two installs using the same name report into one project, and two using different names report into two. Decide your convention before rolling this out to a team. See Configuration for recipes.

​
Install

Run the installer with the argument for your tool. This example traces Claude Code on macOS or Linux:
Replace claude with the tool you want: antigravity, codex, cursor, devin, gemini, copilot, kiro, opencode, or omp. The installer prompts for your license key and project name, then registers the hooks. For the Windows command, the setup prompts in detail, and tracing more than one tool on the same machine, see Installation.

​
Confirm it is working

Use the coding agent as normal, sending at least one prompt. Then open the LLM page in Atatus. Your project should appear within a minute or two of the first trace. If nothing shows up, work through these in order:
  1. Look on the LLM page, not APM. Coding-agent projects never appear under APM, because they have no web transactions to show there.
  2. Check the log for your tool at ~/.atatus/harness/logs/<harness>.log.
  3. Confirm the license key in ~/.atatus/harness/config.json is correct, since an invalid key is rejected at ingest.
  4. Confirm the tool’s own config file still contains the hook entries. The file for each tool is listed in Coding agents.
To see what would be sent without sending it, set ATATUS_DRY_RUN=true and use the tool. Spans are written to the log instead of being transmitted, which is the quickest way to tell a capture problem apart from a connection problem.

​
Read your first trace

Open a trace from the LLM page. You are looking at a tree:
  • A turn at the root, which is one prompt and the response to it. It carries the model name, token counts, and cost.
  • Tool spans beneath it, one per tool the model invoked, such as a file read or a shell command.
Span kinds explains every kind you might see, and Cost and tokens explains the spend figures, including why cached tokens matter so much for coding agents.

​
Monitor your own application instead

LLM monitoring also works for model calls your own code makes, using an Atatus APM agent rather than the coding agent installer. That path is supported in the Python agent. To get started, follow the Python agent overview.

​
Next steps