agent-usage-manager

A local htop-style dashboard for headless AI-agent processes. It groups each agent's process tree, tracks CPU/RAM/GPU and sustained hot/idle/churn/leak states, and provides an allowlisted, token-gated stop control.

Per-node only: AUM is not a scheduler or multi-host control plane. It runs on macOS and Linux; per-process GPU memory is currently NVIDIA/Linux only.

Sanitized capture from a real macOS run. Host, path, and command details are replaced; Apple Silicon does not expose per-process GPU memory.

Quick start

uvx agent-usage-manager

The dashboard opens at http://127.0.0.1:8765. No database or hosted service is required. The bundled config recognizes common tools including Claude Code, Codex, OpenClaw, Hermes, Ollama, vLLM, llama.cpp, Kiro, Aider, and Cline.

Other install options:

pipx install agent-usage-manager
# or, inside a virtual environment:
pip install agent-usage-manager

What it does

• One row per agent. Child inference processes, MCP servers, and helpers are rolled into the owning agent's row. CPU, memory, and GPU values are tree totals.
• Shows change over time. CPU sparklines and hot, idle, churn, and leak? states surface sustained problems rather than one noisy sample.
• Supports alerts. A state transition can call a local command with $AUM_* environment variables. agent-usage-manager test-alert proves the path first.
• Stops the process tree. kill sends SIGTERM and escalates after three seconds; force sends SIGKILL. launchd-supervised jobs get the command that actually stops the service.
• Exposes read-only telemetry. Use agent-usage-manager list --json, GET /api/agents, or Prometheus-compatible GET /metrics.
• Keeps the control local. The default bind is 127.0.0.1; there is no account, remote collector, or central coordinator.

AUM complements process and telemetry tools:

• htop sees processes, but not ownership of a spawned agent tree or agent states.
• Grafana and Prometheus are useful fleet telemetry; AUM can feed them without requiring that stack for a single machine.
• GenAI traces describe sessions, tokens, and cost. AUM covers the host-process side: liveness, resource pressure, restart churn, and local stop control.

Configure agents

Create agents.yaml in the directory where you launch AUM, or pass --config /path/to/agents.yaml:

agents:
  - label: claude-code
    match: "claude(\\s|$|-code)"
    regex: true
  - label: ollama
    match: ollama

protect:
  - uvicorn

ignore:
  - crashpad

Only matched processes are listed or stoppable. protect: keeps a matched process visible but disables stopping it; ignore: removes a false match entirely. Edits hot-reload while the server runs.

Advanced matching, per-instance tmux/launchd labels, alerts, API details, service setup, and troubleshooting are in the operator reference.

Safety boundary

A web page that can stop processes needs a narrow boundary:

• The server re-checks the allowlist before every signal. PID 1, AUM itself, and configured protect: matches are always refused.
• Every stop request requires an X-Kill-Token. AUM creates the token in a mode-0600 state file and never serves it over HTTP.
• Browser requests are checked for DNS-rebinding and cross-origin state changes.
• Non-loopback binds fail closed unless --unsafe-expose is explicit.
• Every stop attempt and refusal is appended to a local action log.
• Common token and key shapes are redacted before command lines reach the UI. Redaction is not a substitute for reviewing screenshots and issue attachments.
• Read-only endpoints still expose process metadata. Keep AUM on a trusted machine or put an authenticated boundary in front of any proxy.

Report vulnerabilities privately through the security policy. Do not open a public issue for a kill-path, authorization, redaction, or package-privacy defect.

Limits

Command-line and API entry points

agent-usage-manager                     # dashboard
agent-usage-manager list                # one-shot table
agent-usage-manager list --json         # scriptable snapshot
agent-usage-manager test-alert          # verify configured alert delivery

The HTTP surface is documented in the operator reference. External controllers should consume the read-only endpoints and own their own scheduling or irreversible decisions.

Project documentation

• Operator reference: configuration, alerts, API, services, proxying, and troubleshooting
• Changelog: current release and unreleased changes
• Security policy: supported versions and private reporting
• Contributing: setup, scope, and privacy requirements
• Architecture and implementation design

If AUM misclassifies a process or behaves differently on macOS or Linux, open a bug report with sanitized output. Never attach a live config, token file, or unreviewed process command line.

AUM is a personal open-source project. Maintenance and support are best-effort.

License

MIT