agent-usage-manager 0.1.3

htop for AI agents — liveness, CPU/mem/GPU usage, and a kill switch for headless agents (openclaw, hermes, ollama, vllm, claude-code).

agent-usage-manager

A tiny, single-file web dashboard for headless AI agents running on a machine — OpenClaw, Hermes, Claude Code, Ollama, vLLM, llama.cpp, or anything you name. It shows which agents are alive and what they're costing you (CPU, memory, GPU), and gives you a kill button per agent.

No database, no auth layer, no dependencies beyond FastAPI + psutil. Runs on macOS and Linux. Meant to be cloned, configured, and run on any node in a fleet.

AGENT          PID    STATUS     CPU %   MEM MB   GPU MB   UPTIME   COMMAND          ┆
openclaw +3    48213  ● running    62.4    1840     7320     2h 11m  openclaw serve … [kill] [force]
claude-code +9 73590  ● running    97.4    7630        —     1h 02m  claude --chann … [kill] [force]
hermes         49001  ● running    18.0     512        —     44m     hermes worker …  [kill] [force]
ollama         50122  ● running     3.1    9210    14080     6h 02m  ollama runner …  [kill] [force]

(+N = child processes rolled up under the agent; CPU/mem/GPU are tree totals.)

A live web UI (auto-refreshing every 3s). The rendered GIF lands here once recorded — see demo.tape.

What it does

• One row per agent. Agents are grouped by process tree — the spawned children of an agent (inference subprocesses, MCP servers, helpers) are rolled up under it with a +N badge instead of cluttering the list as separate rows.
• Liveness — green dot = running, red = zombie/dead. Status column shows the OS state.
• Usage — CPU %, resident memory (MB), GPU memory (MB, NVIDIA only), and uptime, refreshed every 3s. CPU/mem/GPU are tree totals — the agent's true cost including everything it spawned.
• Kill the tree — kill sends SIGTERM to the agent and its children (so spawned helpers don't leak resources), force sends SIGKILL. SIGTERM auto-escalates to SIGKILL after 3s. The confirm dialog tells you how many child processes will stop.

Safety

This is the important part — a web page that can kill processes needs guardrails:

• Allowlist only. Only processes matching a pattern in agents.yaml are ever listed or killable. The kill endpoint re-checks the match server-side before sending any signal, so the dashboard can never be used to kill an arbitrary PID.
• Protected patterns. Anything matching protect: in agents.yaml — plus the monitor's own process and PID 1 — shows a disabled, greyed-out kill button and is refused server-side.
• Secret redaction. Command lines often carry tokens/keys in env vars or flags (FOO_TOKEN=..., --api-key ..., sk-..., ghp_..., JWTs). The command column redacts these to *** before they ever reach the browser — safe to screenshot.
• Bind local by default. It listens on 127.0.0.1. Don't expose it to a network without putting auth in front of it (reverse proxy + basic auth, SSH tunnel, etc.) — it has no built-in authentication.

Quick start

Run it without installing anything (needs uv):

uvx agent-usage-manager
# open http://127.0.0.1:8765

Or install it:

pipx install agent-usage-manager   # or: pip install agent-usage-manager
agent-usage-manager --port 8765

From a clone (for hacking on it):

git clone <this-repo> && cd agent-usage-manager
./run.sh                           # venv + editable install, serves on :8765

It opens the dashboard in your browser automatically. Flags: --host, --port, --config /path/to/agents.yaml, --no-browser (for headless/server use).

Configure which processes are "agents"

Edit agents.yaml:

agents:
  - label: openclaw           # shown as the badge in the UI
    match: openclaw           # case-insensitive substring of the command line
  - label: hermes
    match: hermes
  - label: claude-code
    match: "claude(\\s|$|-code)"
    regex: true               # treat `match` as a regex instead of substring

protect:                      # matched + listed, but never killable
  - uvicorn

ignore:                       # never an agent: not listed, not killable
  - crashpad                  # incidental processes that share a name/bundle
  - shipit                    # path with a real agent (crash handlers,
  - kiro-cli-term             # auto-updaters, integrated-terminal shells, …)

A process matches if the pattern hits its executable basename + first few arguments — deliberately not the whole command line, so a long embedded arg (e.g. a system prompt mentioning "claude") can't misclassify a wrapper. On macOS the outermost .app bundle name is also included, so GUI agents that launch a generically-named binary (Kiro.app → Electron) are still matched by app name.

protect: keeps a matched process listed but refuses to kill it; ignore: drops it from agent classification entirely. Point at a different file with AGENTS_CONFIG=/path/to/agents.yaml.

launchd-supervised agents (macOS)

Some agents run as launchd services (a ~/Library/LaunchAgents/*.plist, or anything started by brew services). If such a job sets KeepAlive, a signal can't stop it: the process dies, launchd immediately respawns it under a new PID, and the dashboard's "kill" looks like it silently failed.

The dashboard detects these (via launchctl list) and marks them with a launchd badge. Instead of dead-end kill/force buttons it shows the command that actually stops the job — click to copy:

launchctl bootout gui/<uid>/<label>            # stop now
launchctl disable gui/<uid>/<label>            # …and don't auto-start at login

The kill endpoint refuses signals for these jobs (HTTP 409) and returns the same guidance, so the API never lies about a kill that won't stick. The message is tailored to the job: KeepAlive jobs are told a signal won't stick at all; RunAtLoad-only jobs are told a signal works now but the job restarts at next login. Limitation: detection runs in your user launchd domain, so root LaunchDaemons (which need a privileged launchctl print system/…) aren't flagged — a known gap, not silently handled.

GPU notes

Per-process GPU memory comes from nvidia-smi when it's on PATH (Linux / NVIDIA). Apple Silicon has no per-process GPU accounting API, so the GPU column stays blank on Macs — CPU and memory are the meaningful resource signals there.

API

• GET /api/agents → { agents: [...], host, cpu_count, ts }
• POST /api/kill/{pid}?force=false → SIGTERM (or SIGKILL with force=true)

Run as a service

Linux (systemd), ~/.config/systemd/user/agent-usage-manager.service:

[Unit]
Description=agent usage manager
[Service]
ExecStart=%h/agent-usage-manager/.venv/bin/uvicorn app:app --port 8765
WorkingDirectory=%h/agent-usage-manager
Restart=on-failure
[Install]
WantedBy=default.target

systemctl --user enable --now agent-usage-manager

Development

git clone https://github.com/minglong51/agent-usage-manager && cd agent-usage-manager
pip install -e ".[dev]"
pytest -q

CI runs the test suite on Linux + macOS (Python 3.9 and 3.12) on every push and PR. Cross-platform note: kill uses psutil's terminate()/kill(), which map to SIGTERM/SIGKILL on POSIX and TerminateProcess on Windows.

License

MIT