Coding agent monitoring traces the AI coding assistants your developers use. Each prompt becomes a trace, showing the model calls it made, the tools it ran, the tokens it consumed, and how long it took. This page covers what is supported and what gets recorded.

​
Supported tools

Ten coding agents are supported. They share one install path and one configuration file, so adding a second tool on the same machine is another one-line command.
Antigravity and Gemini CLI both live under ~/.gemini but read different files, so installing one does not affect the other. GitHub Copilot is the only tool whose config file is project-relative: .github/hooks/hooks.json lives in the repository you run the installer from, so it applies to that project rather than to your whole user account. Every other tool registers once per user.

​
What gets captured

Every trace is a tree, and the tools fall into two shapes depending on how they report their activity. Most tools report the prompt and its model call together, so the turn is the model call: Antigravity, Devin, opencode and omp report model calls separately, so they add a level:
The difference matters when you are looking for cost. In most tools it is on the turn. In Antigravity, Devin, opencode and omp it is on the LLM call spans, and a single trace can hold several of them, because those tools call the model again after each round of tool calls. Add them up for the run.
Whichever shape, each span records the tool name, the duration, and whether it succeeded or failed. Failed tool calls carry an error status, so you can filter for them.

​
What is not captured

By default, the contents of files and command output are not recorded. Prompts and tool names are. This is a deliberate default that differs from recording everything, because tool output is the most likely place for source code and secrets to appear. See Privacy for the full list of what each setting controls and how to change it.

​
Install

One command per tool. On macOS or Linux:
The installer prompts for your Atatus license key and a project name, writes them to ~/.atatus/harness/config.json, and registers the hooks. For the Windows command and the per-tool details, see Installation or the page for your tool.

​
Where the data appears

Coding-agent projects appear on the LLM page only. They have no web transactions, so an APM entry would always be empty.
If you cannot find your project after installing, check the LLM page rather than APM. This is the most common reason an install looks like it failed when it did not.

​
Next steps

  • Installation covers the interactive setup and the Windows path.
  • Configuration lists every setting and environment variable.
  • Privacy explains what is recorded and how to change it.
  • Uninstall covers removing the hooks for one tool or all of them.