What you see
- Every running agent, with its total memory and CPU.
- The whole process tree under it: helpers, shells, MCP servers and anything those started.
- Live figures, so you can watch a leak grow.
Quit or force-kill
Stop a whole agent with everything it started, or a single process such as one stuck MCP server or a dev server left behind by a session that ended hours ago. Quit asks the process to exit. Force kill ends it immediately.
When it helps
- Your Mac's fans are on and you have four agent sessions open.
- A long session has grown to several gigabytes.
- A test runner or dev server the agent started is still running after the agent exited.
- An agent is stuck in a loop and won't respond to Ctrl+C.
The same process list drives the network guard's connection log, so you can see what each of those helpers is connected to as well.
Questions
Which agents are shown?
Any agent Rollcage detects: Claude Code, Codex CLI, Cursor's agent, Gemini CLI, Aider, OpenCode, Goose and Amp by default, plus any process names you add.
Does monitoring use much CPU itself?
No. Rollcage samples the process table every few seconds, which is a cheap system call.
Will force kill lose my work?
It ends the process without letting it finish what it was doing, so an edit in progress may be incomplete. The session transcript up to that point is already on disk, and your project's snapshot is untouched.