Why an agent is hard to see in a process list
A coding agent is never one process. Claude Code runs as node, then starts a dev server to check its work, a headless Chromium to look at the result, a test runner, a tool server or two, and a shell for every command. In Activity Monitor that is a dozen rows called node, Chromium and zsh. None says which chat started it, or whether that chat is still open.
So a session you closed an hour ago can still hold a server on port 3000 and a browser using two gigabytes. With three sessions open, the one waiting for your answer looks the same as the two that are working.
Sessions by name, with their state
Vitals recognises an agent by its command, so Gemini CLI shows as Gemini CLI although it runs in node, and Aider as Aider although it runs in python. The window it runs in does not matter: Claude Code is listed whether it runs in a terminal or inside the Claude app, and Cursor and GitHub Copilot sessions show by name.
Each row gives the project folder, how long the session has run, whether it is working, waiting for you or idle, and the memory and CPU of the whole session with its helpers.
- Recognised: Claude Code, Codex, Gemini CLI, Aider, OpenCode, Cursor, GitHub Copilot, Amp, Kiro and Goose
- Also in the menu bar dropdown’s Agents tab, and as an Agents card on the Overview
- For Claude Code and Codex, the tokens used today and this week
What a session started, and one Quit for all of it
Open a session and everything it started is listed under it: dev servers, headless browsers, test runs and tool servers, each with its own memory and CPU. The same things appear under their project in the Developer tab. A session closed an hour ago is still found, with what it left behind.
Quit a session and everything it started closes with it, so nothing stays on as an orphan. Vitals shows the list and asks first, every time. Force Quit is there for an agent that has stopped answering.
Answer permission requests from the notch
Connect Claude Code, Codex or Gemini CLI in Settings → Integrations → Coding Agents. From then on, when the agent wants to run a command or edit a file, the request comes out of the notch with the command and Allow, Always Allow and Deny. Each project has its own colour, so you know which session you are answering.
Claude Code’s questions arrive there too and are answered with a click. Gemini CLI can only be answered in its own window, so for it the notch tells you it is waiting. When a long turn finishes, a line says so, with what it left running.
Vitals reads the process, its folder and what it started. A connected agent also passes the command or file a request is about, so the notch can show it. Nothing about your code or conversation is kept.
How to use it
- 1
Open the Agents tab
It is in the window and in the menu bar dropdown. Each running session is a row named after its agent, with its project folder.
- 2
Open a session
See the dev server, browser, tests and tool servers it started, each with its memory and CPU.
- 3
Quit a session
Click Quit on it. Vitals lists what will close and asks first.
- 4
Connect an agent
In Settings → Integrations → Coding Agents, click Connect… beside Claude Code, Codex or Gemini CLI to get its requests in the notch.
The details
- Agents recognised
- Claude Code, Codex, Gemini CLI, Aider, OpenCode, Cursor, GitHub Copilot, Amp, Kiro, Goose
- Session states
- Working, waiting for you, or idle, with how long it has run
- Listed under a session
- Dev servers, headless browsers, test runs and tool servers
- Requests in the notch
- Claude Code, Codex and Gemini CLI, once connected. Gemini CLI is answered in its own window
- Notch choices
- Allow, Always Allow, Deny or Answer in Terminal
- Token counts
- Claude Code and Codex, today and this week, from the logs on your Mac
- In Terminal
- vitals agents
Questions
Which coding agents does Vitals recognise?
Claude Code, Codex, Gemini CLI, Aider, OpenCode, Cursor, GitHub Copilot, Amp, Kiro and Goose.
Can Vitals stop an agent by itself?
No. It quits a session only when you ask, and it shows what will close before anything does.
What happens to a dev server when I quit the session?
It closes with the session, together with the browser, the tests and the tool servers the agent started. You see the list first.
Does Vitals read what the agent is working on?
It reads the process, its folder and what it started. A connected agent also tells it the command or file a request is for. Nothing about your code or your conversation is kept, and nothing about what you run leaves the Mac.
Does it count the tokens an agent used?
For Claude Code and Codex, yes: today and this week, where they went and which models took them, read from the logs those tools keep on your Mac.
The same feature on the app's own site: vitalsmac.com/features/agents