Aetower
Developer Preview · v0.54 · macOS
Your Mac runs agents now.
Run it like an operator.
Aetower is the operator console for the machine your agents run on. See friction, energy, money, disk, and repo impact from day one — grow into budgets, incident playback, fleet views, and agents that inspect their own runtime over MCP. Nothing leaves your machine unless you export it.
Why this exists
Agents are workloads you didn't schedule
Coding agents, local models, background inference — they launch processes, spawn helpers, and keep working while you're in a meeting. Your Mac became a small datacenter, and nobody was watching it.
They spend real resources
Battery, watts, dollars, carbon, disk, and thermal headroom. Aetower translates agent activity into those units — per session, per repo, per machine — so the cost of a background run is a number, not a fan noise.
The watcher must be local
You can't observe your own machine through someone else's cloud. One local engine serves the app, the CLI, and an MCP server — so you and your agents read the same live truth, and it never leaves the Mac.
The operator's ladder
Aetower meets you where you are — from your first agent install to running a small fleet. Each rung works on day one; none of them requires the previous.
See what's running and what it costs
Every app, daemon, and agent is one entity with one 0–100 friction score. The loudest thing is at the top, and each one comes with a recommendation in plain English. No dashboards to learn, no metrics to configure.
- Entity grouping: Chrome plus 47 helpers is one row, not fifty.
- The AI-tax panel: GPU, VRAM, and energy per agent — with a warning before the VRAM swap cliff.
- Recommendations that adapt to battery, thermal state, and memory pressure.
Reconstruct what happened
The first time your Mac locks up mid-demo, you'll want yesterday back. Activity keeps seven days of entity-level playback and a timeline that reads like an incident report — so "what was running at 2 PM" is an answerable question.
- Scrub through history: friction, CPU, memory, and disk per entity, at capture cadence.
- Spikes, thermal escalations, restart loops, and agent session costs on one feed.
- Reverse pivots: which process holds this file, port, or socket.
Set budgets, act with guardrails
Once you can see the spend, you'll want limits on it. Per-agent budgets alert when a session runs hot; storage intelligence explains growth and stages safety-tiered cleanup of agent artifacts through the Trash; process actions preview before they touch anything.
- Budgets and alerts per agent, with category toggles and snooze.
- Growth attribution and APFS-aware reclaim — node_modules, target/, DerivedData first, never blind.
- Guarded actions with previews and follow-up verification; persistence and code-signing audits.
Ready your repos, watch every machine
Agents work on repositories and across machines. The Repos tab joins storage pressure, Git state, duplicate clones, per-repo AI spend, and operating-contract readiness; Fleet shows every Mac on your network running Aetower — zero config, no server, no account.
- Check AGENTS.md readiness before pointing an agent at a repo.
- Provider actions: re-run failed GitHub workflows, redeploy Cloudflare Pages.
- Bonjour fleet summaries: CPU, memory pressure, thermal state per machine.
The top of the ladder
Let your agents operate themselves
The app owns a local MCP server on a Unix socket — one engine, no second collector. Your agents query the exact live state you're looking at: 30+ read-only tools for snapshots, history, anomaly explanations, memory breakdowns, storage intelligence, and repo inventory.
After a freeze or a pressure spike, one call to aetower_investigation_bundle hands the agent alerts, findings, recent changes, and history diffs for the window — instead of twenty shell commands.
aetower_investigation_bundle
aetower_explain_anomalies
aetower_reboot_report
aetower_memory_breakdown
aetower_ai_runtime_report
aetower_wakeup_attribution
aetower_diff_snapshots
aetower_recommendations
you › what's straining my Mac right now? agent › calling aetower_top_findings… agent › Ollama (llama3:70b) leads at friction 58 — GPU-bound, 8.1 W. VRAM is at 78%, close to the swap cliff. Consider a smaller quant.
# Registered automatically for Claude when installed; # manual config for any MCP client: { "mcpServers": { "aetower": { "command": "/Applications/Aetower.app/Contents/Helpers/aetower-mcp" } } } # Read-only by default. Exports are privacy-tiered # and redacted. Auto-registration is opt-in.
Plays well with your stack
CLI
The same live data in your shell, scripts, and pipelines.
OTLP → Grafana
Optional metrics export with a documented reference and a ready-made dashboard.
Docker
Local container metadata joined to the entities that own it.
Chau7
Agent sessions, per-repo costs, and runtime state when Chau7 runs.
Chromium debug
Tab-level attribution through compatible debug endpoints.
Shortcuts & shell hooks
Process events can trigger your own automations.
What it inspects
Process pressure
CPU, threads, friction.
Wakeups & energy
Battery, cost, carbon.
Memory
Footprint & leaks.
Storage
Growth, cold data, reclaim.
Code-signing
Unsigned / ad-hoc.
Persistence
Agents, daemons, cron.
Agent impact
Timelines, budgets & MCP.
Repositories
Clones, contracts, spend.
Versus the usual suspects
Built like it preaches
Local-first
Runtime history and diagnostics stay on your Mac unless you explicitly export them.
Conservative
Endpoint Security, telemetry, and automatic MCP registration are all off until you opt in.
Open source
AGPL-3.0, with every binary release paired with its exact corresponding source archive.
Watches itself
The Diagnostics tab turns Aetower's own pipeline into the same observable surface it gives your apps.
Developer Preview
Join the preview deliberately
Aetower is a Developer Preview: signed, notarized, and auto-updating through Sparkle — and moving fast. Surfaces evolve between releases, and feedback steers what hardens first.
It's built for technical people who understand macOS permissions and want a say in what an agent-era operator console becomes. If that's you, you're early on purpose.