Use case

Monitor the AI agents running on your Mac

Claude Code in three terminals, Codex on a branch, Ollama holding a model in unified memory — your Mac runs workloads you didn't schedule, and the built-in tools have no concept of them. Aetower is the operator console for exactly this: what your agents are doing, what it costs, and what to do about it.

What you can see

• Every agent as an entity. Coding agents, local model servers, and inference tools appear as first-class entities with a single 0–100 friction score, not as anonymous rows in a process list.
• Live sessions. With the Chau7 adapter connected, sessions surface with their repository, branch, prompt/approval state, and per-session resource chips — so "which session is the heavy one" is a glance, not an investigation.
• Resource attribution, honestly labeled. Inferred GPU share (macOS has no per-process GPU API, so it's a heuristic and says so), unified-memory pressure, and kernel-measured energy per agent — each with a visible source badge.
• Cost context per repository. The Repos view carries per-repo AI spend, so an agent-heavy project's bill is attached to the project, not lost in a global counter.
• The aftermath. When a session ends or a machine locks up, the Activity timeline narrates what spiked and when — bounded local history you can scrub, entirely on your machine.

Your agents can read it too

The same engine that renders the app serves a read-only local MCP server — 30+ tools for snapshots, history, anomaly explanations, memory breakdowns, storage, and repository state. Point Claude Code or any MCP client at it and ask "what's straining my Mac right now?"; one aetower_investigation_bundle call replaces twenty shell commands. Guarded operator actions exist behind a separate opt-in, and every action stays preview- and approval-gated.

{
  "mcpServers": {
    "aetower": {
      "command": "/Applications/Aetower.app/Contents/Helpers/aetower-mcp"
    }
  }
}

Prefer the shell? aetower top, aetower repos, and aetower storage print the same live state, with --json for pipelines — see the CLI reference.

What Aetower deliberately doesn't claim

• No exact metering. Energy, dollars, and GPU figures are estimates from a consistent model, labeled with source and confidence. Numbers you can act on; not numbers you should bill against.
• No cloud. Everything is local-first: history and diagnostics stay on your Mac unless you export them, and every channel that could leave the machine is opt-in and inspectable under Privacy → Outbound Data.
• No agent babysitting theater. Aetower observes and explains; it doesn't pretend to sandbox or control your agents.

Works with

Anything that runs as a process is visible. Session-level detail comes from the Chau7 adapter; local model servers such as Ollama and other inference tools are recognized as AI runtimes; and any MCP-capable agent — Claude Code, Codex, and friends — can query Aetower directly.

Questions this page answers

How do I monitor Claude Code's resource usage on a Mac?

Aetower shows each agent as an entity with a friction score, kernel-measured energy, and — with the Chau7 adapter — per-session repo, branch, and resource detail.

How much energy do local LLMs like Ollama use?

Aetower reports kernel-measured per-entity energy and inferred GPU share for local model servers, labeled as estimates with their source — not exact metering.

Can my AI agent check what is straining my Mac?

Yes — Aetower's read-only local MCP server exposes 45 tools; one aetower_top_findings or aetower_investigation_bundle call gives an agent the ranked answer.