Building / 03
Cracked
My Macs run more agents, tabs and local models than is reasonable, and Activity Monitor only tells me what's happening right now. So I built Cracked, a menu bar app that says when the Mac is struggling and why: which of my Paseo and Claude Code agents need me, which local models are sitting in RAM, and what's been eating it for days. It's Swift 6 and SwiftUI, and everything stays on the Mac.
- Role
- Solo build
- When
- Feb 2026 to now
- Team
- Solo, plus one outside contributor
- Status
- in progress
01 / The build6 figures
Fig. 02
I wired it into my coding agents, from Paseo and Claude Code.
Every agent shows up as working, done or needs input, with what it's waiting on. The blocked and finished ones get a count next to the icon, and a click opens the agent in Paseo or jumps to its exact terminal tab.
Field notes
Paseo needs no setup: its daemon rewrites a small JSON record per agent in ~/.paseo/agents whenever the agent's state changes, and Cracked reads those, re-decoding only the files that changed. A permission prompt reads "Waiting for permission", an error shows Paseo's last error, and a click opens paseo://h/<server>/agent/<agent> in the Paseo app. Claude Code sessions in a terminal report through hooks (SessionStart, UserPromptSubmit, Notification, Stop and SessionEnd) or the files Claude Code keeps in ~/.claude, and a session Paseo is running is listed once, under Paseo, where you answer it.
GET 127.0.0.1:11434/api/ps every 5s, while Ollama runs
POST /api/generate
{"model": "gemma4:26b", "keep_alive": 0}
what `ollama stop` sends
Fig. 03
It shows the local models hiding in RAM, and frees them in one click.
Models loaded in Ollama take tens of gigabytes and barely show up in process stats, because their GPU memory is wired. Cracked asks Ollama what's loaded, shows how much of the Mac's RAM they hold, and Unload drops one right away.
Field notes
It polls Ollama's /api/ps every five seconds, only while Ollama is running, and only on this Mac. Unload sends what `ollama stop` sends: a generate request with keep_alive set to 0 and no prompt, then /api/embed for embedding models that refuse it. Ollama reloads the model on next use. The section header turns yellow or red with macOS's own memory pressure, and when the Mac is under load, new agent sessions are told which models are loaded and asked not to load more.
Fig. 04
I built one icon that tells me how my Mac is doing.
The gauge in the menu bar stays calm until an app goes over 100% CPU, load passes 20 or memory pressure turns critical. Click it for everything at once: who needs you, the models, the apps eating your cores, and the ones worth a restart.
Field notes
Worth a Restart isn't just "apps that are old". Messages and Slack are meant to run forever, so an app only shows up when it has been up for days and is holding a lot of memory across all its helper processes, like a browser with 170 renderers sitting on 65 GB. Restart quits it gracefully and relaunches it, and you can snooze a row for a week.
Chill
Everything's fine. Probably.
Warning
An app over your CPU threshold, load above 8, or memory pressure
Cracked
An app above 100% CPU, load above 20, or critical memory pressure
Fig. 05
It has three moods, and it only gets loud when it should.
Chill, warning and cracked, with thresholds you can tune in Settings.
Fig. 06
It reads the system the boring way, off the main thread, and talks back to agents.
Every few seconds (five by default) Cracked polls ps, sysctl and vm_stat in the background. Snapshots go into SQLite through GRDB and get downsampled: per minute for a day, per hour for a month, then daily.
Field notes
An early version hung the app by running shell commands on the main thread, and another deadlocked on a full pipe. Both fixes are in the history. Cracked is also a small hand-written MCP server, from the same binary, so agents can ask for system_status or list_agent_sessions, and with my approval restart an app. When the Mac is struggling, new sessions are told so ("load 46 on 16 cores · Xcode 140%") and run builds one at a time instead of piling on. Records & Stats keeps 24-hour charts, all-time records and a Hall of Shame of the worst sustained CPU.
Fig. 07
Next, my Macs should see each other.
The plan is a fleet over Tailscale, so the Mac mini and both MacBook Pros share state and an agent can pick the least loaded Mac. Then share cards, an opt-in leaderboard by chip, and llama.cpp's llama-server next to Ollama, with its own Unload or Stop.
Field notes
The leaderboard only ever gets numbers, a chip class and a fixed list of app names, kept as a separate type in code so nothing private can leak into it. Fleet and leaderboard both stay off until you turn them on; with them off, Cracked makes no network calls beyond local model runtimes.
02 / Creditswho made it
03 / On the bench