What it does

agent-top gives people who keep several coding agents running at once a single terminal screen to watch them on. It reads the transcripts that each agent harness already writes, together with the operating system's process table, and shows all the sessions it can see in a single view. For every session it shows whether the agent is working or waiting for you, what it has cost so far at list price, and which helper processes are still around. MCP server processes whose agent has exited are highlighted in red, so leaks are easy to spot. It is a single static binary for macOS and Linux that needs no daemon or configuration, and it does not write to, signal or kill any agent. The site states that your data stays on the machine. Cost reports can be broken down by harness, model, project or day, and the whole screen can be saved as a single JSON snapshot for bug reports.

Capabilities

Live view of Claude Code, Codex, Gemini CLI and OpenCode sessions in one place
Shows which agents are busy and which are waiting on you
Token and dollar cost counted from each harness's own usage records
Flags MCP server processes left behind after their agent has gone
Cost reports by harness, model, project or day
Saves the screen as a JSON snapshot for replay and bug reports

When to use it

You run several coding agents in separate windows and want one overview
You want to see what each agent session has cost so far
You suspect orphaned MCP server processes are piling up