🔥 Your Mac is slow. AI Agent finds the culprit.

English · 한국어 · 简体中文 · 繁體中文 · 日本語 · Español · Deutsch · Français · Português · Русский

Your Mac chokes → Claude names the exact process eating it → one click kills it. Nothing happens until you click — so there's no risk to your Mac.

Every hour, your Mac's load goes to Claude. It ranks what's actually hogging your CPU/RAM, writes the exact fix, and drops it in your menu bar — worst-first, color-coded, one click away. And before anything runs, a second Claude pass has to clear the command as SAFE.

Mac Optimizing Looper is a macOS menu-bar app (no Dock icon) that runs a continuous observe → ask the model → advise → (optionally) act loop on top of your local LLM CLI.

⬇ Install · See it work ↓

Activity Monitor shows you 200 rows and zero answers. This shows you the one command that fixes it — and why.

How it works

One cycle:

⏱  timer fires (default: every 1h, slider 10m … 36h)
→  collect: CPU/MEM + mac-optimizer snapshot (+ optional sustained sample)
→  claude -p   (analysis pass, --effort max)
→  claude -p   (format pass → ranked JSON suggestions)

The menu bar shows the count. The dropdown is ranked worst-first: 🔴 critical → 🟡 warning → 🟢 hygiene. Each row expands into Copy · Show in Terminal · Review with Claude · Run Command Now.

The safety gate — why it won't nuke your Mac

"Run Command Now" is the only path that executes anything, and it's gated end to end:

click ▸ Run Command Now   ($ kill 8123)
→  claude -p   classifies → RISK: SAFE
→  background run   (sudo → GUI password prompt, because there's no TTY)
→  ✅ notification → click → full stdout/stderr window
→  suggestion marked ✓ done

Anything not classified SAFE (including unknown) pops a confirmation dialog defaulting to Cancel. The advice itself is inert data; the model can never make the app run a thing. Enforced by GuardrailTests.

Mac Optimizing Looper vs the usual suspects

Install

Needs the claude CLI on your PATH. macOS 13+.

brew install --cask kargnas/tap/mac-optimizing-looper

The cask + DMG go live after the first signed release. The pipeline is wired and waiting on signing secrets — see docs/release-setup.md. Until then, build from source:

git clone https://github.com/kargnas/mac-optimizing-looper
cd mac-optimizing-looper
bash script/build_and_run.sh run     # builds the .app, codesigns ad-hoc, launches

Run the bundle, not the bare binary — UNUserNotificationCenter needs a real bundle id (as.kargn.MacOptimizingLooper).

Make it yours

Switch Provider / Model / Effort right from the menu bar — no need to open Settings. Each has a Default option that auto-picks for you: Claude as the backend, a balanced model, and one notch below the top reasoning level. A fresh install just works, and you can still pin an exact value any time. While an analysis runs, the menu bar shows which backend is working. claude is the default backend; codex is also supported (one schema-constrained pass, no separate format step). Settings holds the rest — Fast Mode, analysis interval, terminal app, and Language (which sets both the UI and the language the model analyzes in). Models and reasoning levels are read live from each CLI.

FAQ

Does it ever run anything by itself? No. Advice is inert data. The single execution path is the "Run Command Now" button, on your click — enforced by GuardrailTests.

Is it safe to hit "Run"? Every command goes through a second Claude pass. Anything not clearly SAFE (including unknown) pops a confirm dialog defaulting to Cancel. sudo routes through the macOS GUI password prompt.

Does my data leave my Mac? Only the live metrics + process table, sent to Anthropic via your own claude CLI (or OpenAI via codex). The app adds zero telemetry.

What does it cost? Nothing beyond your existing claude / codex CLI usage. The app is free and MIT-licensed.

No claude CLI installed? Then no advice — it surfaces the error instead of guessing.

You are a macOS performance analyst.
Given live metrics + a process table, identify the ACTUAL remediation
command (kill / killall / unload) for each real problem — never an
inspection command (no pgrep / ps / top).

MUST:     rank by severity; prefer graceful `kill <pid>` over `kill -9`.
MUST:     return a null command when no command-line action applies.
MUST NOT: claim anything was executed — the app never auto-runs.

timer → collect → claude analyze → rank suggestions
                                       │
                 user picks an action ─┼─ Copy / Show in Terminal → no execution
                                       ├─ Review with Claude       → interactive claude session
                                       └─ Run Command Now
                                              → claude risk-check
                                                   ├─ SAFE → run → notify → ✓
                                                   └─ else → confirm (default Cancel)

MIT licensed. Built for people who'd rather know why their Mac is slow than reboot and hope.