MacPop pre-release testing: the results so far
What it cleaned, what it left alone, and the memory and CPU cost of running it.
Pre-release results · Public downloads are not open yet
MacPop finds caches, build files, and background processes left by developer and AI tools. These pre-release experiments look at what it can clear, what stays protected, and the cost of running the app.
The cleanup is only useful if it leaves current work alone. The other measurements cover scanning speed and background resource use.
What it cleaned and what it left alone
For the November 14 cleanup run, I prepared about 4.6 GB of test files. Some were disposable caches and build output. Others represented work that should stay untouched.
| Situation | Observed result | What that means |
|---|---|---|
| Two selected rebuildable-cache items | 357 MB removed after confirmation | The selected cleanup freed space; it did not remove the whole 4.6 GB dataset. |
| A project with a running development server | Left alone | MacPop recognized that the project was in use. |
| Work with uncommitted changes, and files Git did not ignore | Protected | These were treated as work to keep. |
| Agent conversations | Excluded from cleanup | Conversation history was not offered for removal. |
| Recovery records could not be saved | Cleanup refused | The chosen cache stayed untouched when the recovery guard could not proceed. |
The point is to review a cleanup and understand its effect. Removing a cache can mean downloading or rebuilding it later. More free space is useful, but the next build still matters.
A separate 20-minute scripted workload exercised background cleanup. MacPop stopped 42 leftover helper processes after their agent sessions ended, with a median delay of about 23 seconds. The workload's active sessions and builds were left running. These were deliberately created workloads, so they show behavior under those conditions rather than a promise about every app or process on a Mac.
How long scanning took on my Mac
These measurements were recorded on November 7 on an Apple M4 Mac with 16 GB of memory, using a development build. The idle sample lasted about two minutes, with the app's windows closed.
| Measurement | Recorded result | How to read it |
|---|---|---|
| Space Map of the Developer folder | 1.86 seconds for 37.1 GB and 568,481 files | Time to map this specific folder. |
| Broader source scan | 24.7–25.3 seconds across two runs | A broader check takes longer than mapping a single folder. Desktop, Documents, and Downloads stayed closed. |
| MacPop memory while idle | 78 MB | Memory used by the app itself during this short sample. |
| MacPop CPU while idle | 0.90% of one CPU core, on average | CPU used while the app sat in the background. |
Scanning shows where space is used. It does not mean that amount was deleted: the 37.1 GB folder map and the 357 MB cleanup were separate operations, recorded on different dates.
With and without the scheduler
I also compared the memory scheduler switched off and on in a scaled simulation. Four stand-in agent workloads ran for eight minutes in each condition, using the same workload and random seed.
These are simulated jobs. A paired comparison of real project build times is still to come.
Eight minutes per condition · scaled, synthetic workload
| Simulated workload result | Scheduler off | Scheduler on | Difference |
|---|---|---|---|
| Average memory held by the agent workloads | 698 MB | 429 MB | 269 MB less |
| Peak memory held by those workloads | 736 MB | 532 MB | 204 MB less |
| Work steps finished in eight minutes | 24 | 37 | 13 more completed |
| Average time from requesting a step to finishing it | 66.4 seconds | 44.2 seconds | 22.2 seconds less, including queue time |
The scheduler spread the work out instead of letting every heavy step compete at once. In this model, that reduced the memory held by the workloads and let more work finish within the run.
There was a tradeoff. Some steps waited before starting; the longest wait was about 106 seconds. MacPop's own average CPU use rose from 0.48% to 0.84% of one core. The simulation checked that the terminal session in active use was not parked.
The model used a 600 MB agent budget and a 100 MB reserve, representing 6 GB and 1 GB at one-tenth scale. It deliberately slowed jobs to 80% speed at warning pressure and 40% at critical pressure, with further slowdown beyond capacity. The faster completion times depend on those assumptions. They do not establish how much faster a real npm install or Xcode build would be.
What I am checking before release
The next step is a repeated comparison using real projects, with Autopilot off and on. That needs to include the cost of rebuilding a cache after cleanup, the effect on active work, and the app's resource use over longer sessions.
MacPop remains in pre-release testing while I work through those checks. If you want to hear when public downloads open, join the launch list. Confirmed subscribers will receive a one-use coupon for the $4.99 launch price, before applicable tax.