clawtop

htop for OpenClaw. See where your gateway's memory goes, per agent run; stop the runaway ones safely; keep your sessions tidy.

_{clawtop top on sample data (kills simulated). Also: a 1:43 video walkthrough of the command-line tools.}

Documentation: https://khiladisngh.github.io/clawtop/

Why

Many OpenClaw runtimes (CLI backends such as Claude CLI or Gemini CLI, and ACP harnesses) start a separate process for every agent run, plus the MCP servers and helpers that run uses. In the author's setup each such stack is roughly 0.5 to 0.7 GB. A burst of runs, or a few that hang on a tool call, can push the gateway past the VM's RAM; the event loop then stalls in swap and restarts fail. clawtop makes that visible and fixable in one command. It is not tied to one runtime: runs are recognised by structure, not by program name.

• Memory per run, measured so the numbers add up to the gateway's cgroup total.
• Real idleness where the runtime exposes it (Claude Code's transcript today), so a busy long-lived run is not flagged and a stuck one is. Where it doesn't, idle time is shown as -, never guessed.
• Safe kills. Only processes the gateway spawned can be selected, whole stacks die together (SIGTERM, then SIGKILL), and PID reuse is guarded.
• Sessions through the supported openclaw sessions CLI: list, tree (subagents under their parent, with titles), archive, prune.
• History: a per-minute sampler and clawtop history to see what a spike looked like.
• Works from Windows: it forwards into your WSL distro.

Install

On the Linux or WSL machine that runs the gateway (needs uv and Python 3.11+):

uv tool install git+https://github.com/khiladisngh/clawtop
clawtop
clawtop sampler install     # optional: per-minute memory history via a systemd user timer

On a Windows host, install it there too; every command is forwarded into WSL (your default WSL distro and user; override with CLAWTOP_WSL_DISTRO and CLAWTOP_WSL_USER). See Install.

Usage

kill and sessions archive|prune ask for confirmation; --yes skips it and --dry-run previews. prune only touches subagent sessions unless you pass --kind, because channel and main sessions hold conversation context. Full options: CLI reference.

Try it without a gateway

scripts/demo.py runs the real CLI against made-up sample data (the video above uses it); kill and archive are simulated:

git clone https://github.com/khiladisngh/clawtop && cd clawtop && uv sync
uv run python scripts/demo.py sessions tree
CLAWTOP_DEMO_ANSWERS="stale;y" uv run python scripts/demo.py kill

Keep the gateway from running out of memory

clawtop shows problems; systemd limits and a concurrency cap stop them taking the VM down. See Harden the gateway.

Limits

• Linux only (reads /proc and cgroup v2); on Windows it needs WSL.
• It looks for the systemd user unit openclaw-gateway.
• Runtimes that run inside the gateway (openclaw, codex, copilot, muse) have no process per run, so there is no stack to list; their memory is part of the gateway's "main".
• Verified on Claude CLI only. Other process-per-run runtimes (Gemini CLI, ACP harnesses) are recognised by the same structural rules, but I have not seen them on a real gateway. Please report what you see. See Concepts.
• A run stack cannot be mapped to an OpenClaw session key (the gateway does not expose it), so clawtop shows the runtime and agent instead of guessing.
• Runs are found by agent workspace (from your OpenClaw config), so the config must declare them.

Development

uv sync
uv run ruff check . && uv run ruff format --check .
uv run pyright
uv run pytest --cov
uv run zensical serve     # docs at http://localhost:8000

The /proc parser tests build a fake /proc tree and need Linux; they are skipped on Windows and run in CI. See CONTRIBUTING.md.

License

MIT