Monitor Claude Code, Codex and other AI coding agent sessions on your Mac

Why watch your coding agents?

An AI coding agent doesn't just write code. It starts things: a dev server to check its work, a test watcher, a TypeScript build, a Python script that never exits. Close the terminal tab and many of them keep running. After a day with two or three agents, it's common to find a gigabyte of memory and half a dozen ports held by processes nobody remembers starting.

Activity Monitor shows them as a flat list of node and python entries with no hint of which agent or project they belong to. Dev & Agents groups them back into sessions, so you can see that Claude Code in storefront-web is responsible for four processes, 1.18 GB and port 3000.

Open Dev & Agents under Your Mac in the sidebar (or Go › Dev & Agents); Agents is the first tab. The other tabs, for dev servers, Docker and build folders, are covered in Manage dev servers and ports.

Which AI agents are detected

OneMacTools recognises these command-line agents by their program name or how they were launched:

It doesn't matter where they were started: Terminal, iTerm, VS Code's terminal, Cursor, or the Claude desktop app. The from chip on each session names the app.

When none is running you'll see No AI agents running. Sessions appear while they run and disappear when they end.

What each session card shows

• The agent and project: Claude Code storefront-web. The project is the nearest folder with a package.json, .git, Cargo.toml or similar project file; the full path is shown underneath.
• from Terminal (or the app that started it).
• Port chips like :3000, for up to three ports held by the agent or anything it started.
• running 2 h 12 min · 4 processes: how long since it started, and the agent plus what it launched.
• Memory in large type and CPU underneath, for the whole session added together.
• Started N processes: expand it to see each command, its process ID and its memory (up to 40).

Sessions are listed newest first. Memory comes from macOS's own accounting; CPU is the share of one core used since the last reading, so a session at 150% is using one and a half cores.

Sub-agents and hosts

When an agent launches another of the same kind in the same folder (a sub-agent), its processes are counted in the parent's session rather than as a separate card. An agent process that only hosts sessions in other folders, running from your home folder or /, isn't counted as a session itself, so the desktop apps that run Claude Code for you don't create a phantom card.

How to stop a Claude Code or Codex session

Sometimes you want an agent gone, with everything it started: it's looping, it's holding a port you need, or you've finished and the terminal is long closed.

1. Click Stop… on the session card.
2. Read the confirmation, for example Stop this Claude Code session? with Stops the agent and the 3 processes it started, freeing port 3000. Unsaved work in the session may be lost.
3. Click Stop. A green line confirms how many processes stopped.

OneMacTools asks each process to quit, starting with the newest children and ending with the agent, the same polite signal as pressing Ctrl-C. Files the agent has already written stay as they are, and Claude Code and Codex keep their transcripts, so you can usually resume later. Only your own processes are touched, and each is checked to be the same program it listed.

To stop just one server an agent started and leave the agent running, use Stop… on that server in the Dev Servers tab instead.

Alerts when an agent's process runs away

OneMacTools also watches what agents start, every 30 seconds, even with the screen closed. When a process uses a full core for 10 minutes, grows by a gigabyte in an hour, or sits idle on a port for a day, you get a notification with a Stop button. The agent itself is never stopped by these alerts. Set or turn off automatic stopping in Settings › Runaway Processes (settings reference).

Codex and Claude Code usage limits

Above the sessions, Usage limits shows how close you are to being cut off. It's read from the agents' own session files on your Mac, as of their last session, and only the lines about limits are read, never your conversations.

Codex: percent used

Codex records its limits as it runs, so you get real numbers: one row for the 5-hour limit and one for the Weekly limit, each with a bar, 38% used and when it resets. The bar is green, orange from 70% and red from 90%.

Once a reset time has passed, the old figure no longer says anything, so the row reads Reset with when, and the next session will show how much is used.

Claude Code: when a limit was hit

Claude Code doesn't record how much of its limit is left, only when you hit one and when it resets. So its row shows one of:

• Hit the 5-hour limit · resets 1pm: you're limited now, until that time.
• Not limited now · last hit the weekly limit yesterday: you were limited before, but it has reset.
• No limit hit in recent sessions. Claude Code doesn't record how much of its limit is left, only when one is hit.

If Codex is installed but hasn't recorded a limit yet, its row says No limit recorded yet; it shows after the next session. Click Refresh in the toolbar to read the limits again.

How this differs from Activity Monitor and ps

For the whole Mac, including apps and system processes, use Mac Health. Dev & Agents is deliberately narrower: your own developer processes, nothing else.

Agents checking on agents (MCP)

Your agents can see this too. Through the OneMacTools MCP server, Claude Code, Codex, Cursor and others can list agent sessions with what they started, check usage limits, and see what's on a port before starting a server of their own. Anything that changes your Mac still needs your approval (assistant permissions).

Dev & Agents is part of OneMacTools; see pricing for what each plan includes, or download the app to try it.

Questions

Related docs