If you’ve ever had three Claude Code sessions open across three projects and lost track of which one was about to hit a context limit, this post is for you. The tool is called abtop, and it’s exactly what the name suggests — htop, but for AI coding agents.
What it shows
One terminal window, every running agent at a glance. For each Claude Code and Codex CLI session it tracks:
- Tokens used and context window fill % — per-session bars with warnings when one is getting close to full. This is the headline feature for me.
- Rate limit status — your quota in real-time, so you stop guessing whether you’ve been throttled.
- Child processes and open ports — if an agent spawned a dev server and forgot to kill it, you’ll see the orphan port and can clean it up with one key.
It also catches subagents, git status, and the current task each session is on. All read-only — no API keys, no auth.
Why this matters
The dumb thing about running multiple agents is that you can’t actually see them. Each one lives in its own terminal pane, you have to switch around to check status, and there’s no system-level view of who’s burning tokens or about to compact. abtop is the missing system view. One screen, every session, live.
Installing it
macOS or Linux, one line:
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/graykode/abtop/releases/latest/download/abtop-installer.sh | sh
Or via Cargo if you have Rust set up:
cargo install abtop
Then just run abtop. Recommended terminal size is 120x40 or larger.
Windows users — it needs Unix tools (ps, lsof), so run it inside WSL.
Bonus: tmux integration
If you already work in tmux, this gets better. abtop discovers sessions across panes, and pressing Enter on a row jumps you straight to the terminal running that agent. So the flow is: scan abtop → see the one at 87% context → Enter → you’re there, ready to wrap up or compact.
That’s it. Tiny tool, dead-simple install, and once it’s in your stack you’ll wonder how you managed running parallel agents without it.
Links:
