The gap was never the automation

Lots of programs already tidy your Mac while you are not watching. What none of them do is trust themselves more when they have waited longer.12

Here is why that matters. If a program checks every fifteen minutes and gets one thing wrong, it has thrown away something you needed and you find out much later. If it only wakes once a week, it can afford to stop and ask you first. So the safe rule is not "check often" or "check rarely"; it is that the more often it runs, the less it should be allowed to touch on its own.

Unattended action is not the gap. It already ships: a ten-minute launch agent that kills processes without asking,2 a rule engine that has acted on files for about twenty years,3 and until July, a product that cleaned enrolled Macs on a schedule and forced the job through if the user skipped it.1

What none of the three does is vary its permission by cadence. Each has one permitted action set and one timer. The unoccupied position is the coupling: a fifteen minute job allowed only what it can prove is regenerable, a weekly one allowed to propose deleting a git worktree; because the cost of being wrong is unchanged by how recently you last looked.

Scheduled unattended deletion on macOS has at least three shipping precedents: cc-reaper's 600-second LaunchAgent,2 Hazel's rule engine,3 and CleanMyMac Business's 7/14/30-day Smart Scan automation until 2026-07-14.1 None couples the permitted action set to the schedule period.

The coupling is what is unoccupied, and the argument for it is a cost asymmetry rather than a convenience one: deletion risk is a property of the target's reproducibility, and it does not decay with observation frequency. A high-frequency tier therefore has to be narrower, not merely faster.

Limit. This is assembled from three examples, not an exhaustive survey. The absence of a fourth is weaker evidence than the presence of these three.

One big company sold a version of this to businesses. It cleaned the Macs at a company on a timer, and if someone kept putting it off, it went ahead anyway. They stopped selling it in July, and did not say why.1

The research done for this page pointed at that product to prove the idea works. The page it pointed at now says the product was cancelled.4

CleanMyMac Business was the single commercial product shipping scheduled unattended macOS cleaning: Smart Scan automation on 7, 14 or 30-day intervals, with forced execution if a user skipped it. MacPaw withdrew it effective 14 July 2026.1

A commissioned research backend rated that product its strongest evidence for the central question, at "Medium to High Confidence", and cited its documentation more than a dozen times as proof the model was commercially viable. Dereferencing that URL returns the discontinuation notice.4

What this does not establish. MacPaw gave no reason beyond focusing on other products. Nothing here shows the automation failed, or that it caused the exit. What it does mean is that the product can no longer be cited as evidence that scheduled unattended cleaning sells.

CleanMyMac Business Smart Scan automation: admin-configured tasks with a start datetime, repeat interval of 7, 14 or 30 days, a forced-execution day and time on skip, and a configurable reminder count. Cleaned system log files, broken login items, user cache files and language files across enrolled devices. Schedule parameters were immutable after creation. Discontinued effective 2026-07-14, announced 2026-07-29.1

The failure mode this exposes is general: a live-resolving URL is not a live claim. Citation verification proves dereferenceability, not that the content still supports the proposition attached to it. The panel's citation resolved perfectly.4

Limit. Causation is unestablished. A B2B withdrawal is equally consistent with go-to-market economics, support cost, or portfolio strategy. Both the "demand failed" and "distribution failed" readings survive this evidence intact.

One project already does this every ten minutes. It watches for programs that have lost the thing that started them and are now stuck burning the machine's power.2

The clever part is the waiting. It will not act on one look, because one look cannot tell a program that is working hard from a program that is broken; both look identical. So it waits for an hour of bad behaviour first.

At least six people built this same tool without knowing about each other. Almost nobody found any of them.5

cc-reaper runs a launch agent every ten minutes, finds processes whose parent has died, and kills them; unattended, no prompt. It requires 80 percent average CPU sustained for at least an hour before it treats one as a runaway, which is a sustained test rather than a snapshot.2

It also scopes the kill by identity, never by name: group cleanup targets only groups whose leader is a session it owns, and it "never matches by group membership". A separate team hit the same wall and built a registry proving a process belongs to their own config path before killing it.6 Two teams, no contact, same rule.

Six independent implementations of this idea exist. None has more than fifteen stars.5 The idea is not rare. Finding it is.

cc-reaper: LaunchAgent at 600s detecting PPID=1 orphans; runaway trigger requires CC_RUNAWAY_CPU ≥ 80 and CC_RUNAWAY_MIN ≥ 60 minutes elapsed. Companion janitors at hourly (disk) and Sunday 04:00 (weekly clean). It rejects TTY filtering on measured grounds; in SSH, Docker and tmux every process reports TTY=?. Ancestor guard walks $$ → PID 1 and never kills a forebear.2

Its protection list carries a deliberate override: Chrome, Spotlight and shared MCP servers are protected, but "a process pinned at high CPU for hours is broken, regardless of category". Cloudflare's MCP server is named as explicitly not protected because it orphans and pins CPU.2

Convergent design worth noting: ownership proof by config path (Empryo) and by process-group leader identity (cc-reaper), both rejecting pattern matching, which cannot distinguish your runaway from another session's healthy worker.6

Limits. cc-reaper is at 15 stars; established as existing and well-designed, not as adopted. The six-implementation count comes from a single query shape on one day and is a floor, not a census.5

When these helper programs are forgotten, they do not just sit there quietly. One of them kept growing until it had taken about a hundred gigabytes; enough to fill a laptop.7 On another machine, forgotten helpers crashed the computer outright.8

This is no longer a housekeeping problem. An orphaned tool server reached roughly 100 GB on an M4 Max.7 On 16 GB machines, orphaned language-server and shell instances grew about 1.25 GB each per 35 minutes and produced kernel panics with 50 GB of swap swell.8 One team's telemetry found a single machine holding 127 orphaned processes.6

That is the honest case for acting rather than advising. A tool that notices this and then writes you a recommendation has not solved it.

Three primary reports: oraios/serena #1367, ~100 GB RSS from an orphaned MCP server on an M4 Max;7 anomalyco/opencode #12687, ~1.25 GB per instance per 35 minutes from unbounded output concatenation (O(n²)), causing kernel panics and ~50 GB swap swell at 16 GB unified memory;8 Empryo telemetry, 127 orphaned processes holding 725 MB on one host.6

Limit, and it is a real one. These are issue-tracker reports and a changelog. Severe cases are what get filed, so all three are upper-tail by selection. Nothing here establishes a typical figure, and none should be read as one.

To describe what people really use, the research picked projects that almost nobody has starred. Two of them had none at all; not one person had marked them as interesting.910

Meanwhile the one tool here that a lot of people genuinely do use, it described at the wrong price, the wrong size, and under the wrong licence.1213

All of that was free to check. It took seconds.

Both paid backends built their picture of the free market on repositories with 0, 0, 9, 11 and 15 stars. One with zero was cited about twenty times by one backend as its free-market evidence base, and named by the other as evidence of "actual adoption of AI in this space".10 A nine-star repository was called "the strongest direct competitor".9

The same panel got Mole wrong three ways: $19, not $9; 63,562 stars, not 59,100 or 28,500; GPL-3.0, not MIT.1213 The price was on the vendor's own front page. The backend sourced a Medium post instead.11

A vendor page, one API call and one URL fetch corrected every load-bearing number. The research cost about $29.70. The corrections cost nothing.

Star count is a weak proxy for use. It is the metric here because it is the proxy those reports leaned on implicitly, and no backend queried the API that returns it.910

tw93/Mole: vendor-stated $19 one-time for 2 Macs against a reported $9.00;12 63,562 stars and GPL-3.0 against reported 59,100 / 28,500 and MIT.13 Four of four load-bearing corrections resolved against first-party artifacts at zero marginal cost.

The reading that survives: the panel's value was in disagreeing enough to show where to look, not in the figures it returned. Three backends produced three incompatible answers on the central question, and that disagreement is the only reason anyone went and read the artifact.

Limit. Prices change. The $9 figure may have been correct at some earlier date; it was not correct on the day it was reported.

People mind much less about a program deleting something the computer can rebuild by itself than about it touching a folder where they keep things. When one cleaning program started removing files from the Downloads folder, the complaint was not that it was badly built; it was that lots of people keep things there.14

Most tools in this area have decided the safest thing is to tell you what to do and never do it themselves.17 Whether the safe part is big enough for anyone to make a living from, nobody here can tell you. The people who looked at this do not agree, and the one company that tried has stopped.

Tolerance tracks the target, not the tool. Caches and build output are forgiven because the machine can rebuild them; Downloads and Documents are not. The CCleaner moderator's objection was precisely that; not that the feature was broken, but that "many people use Downloads as long-term storage".14

The surrounding sentiment is harder than that. A senior Apple Support contributor advises never installing this class of software at all.15 Sensei sells against automation in its own marketing testimonial.16 And the prevailing engineering posture is refusal: one plugin documents its process skills as "pure diagnostic", stating they "never kill processes on their own" and "merely suggest kill commands".17

This is the question the page does not close. Whether the forgiven set is commercially large enough is genuinely unresolved, and the CleanMyMac Business withdrawal is consistent with either answer.1

The tolerated envelope is defined by the reproducibility of the target, not by the sophistication of the gate in front of it. A perfect verification gate on ~/Downloads is still outside the envelope; a crude one on DerivedData is inside it.14

The stronger version of the objection is not about risk at all. Howard Oakley argues macOS already empties the caches in question and asks why a second application should need to exist; which, if right, makes part of this category redundant rather than dangerous.18

Unresolved, and stated as unresolved. This corpus cannot settle whether the envelope supports a business. The single commercial datapoint terminated without a stated cause and is consistent with both the demand-failure and the go-to-market readings. Nothing in the research, and nothing found on this machine, breaks that tie.

The version of these apps you get from Apple's store has had its most useful parts taken out; not because they were dangerous, but because the store's rules do not allow them. The company that makes the best-known one publishes the list itself.19

The App Store build of the category leader cannot clean system logs or caches, touch Xcode simulators, manage launch agents, or monitor for malware in real time. MacPaw publishes that list on its own support site and is explicit that the exclusions do not mean the features are harmful.19

Underneath that, the paid and free tools have converged on the same targets. Pricing runs from $9.99 once for five Macs20 to $65.99 a year,21 with Sensei at $29/yr or $59 once22 and Mole at $19 once.12 The real lever on perpetual buyers is not a feature gate at all: a one-time licence excludes major-version upgrades.23

Per MacPaw's own knowledge base, updated 2024-10-16, sandboxing plus review guidelines remove system log and cache cleaning, Xcode simulators, language files, iOS device backups, launch-agent enable/disable, menu-bar hardware temperatures and real-time malware monitoring from the App Store build.19 A tool intending unattended launchd operation is therefore off that channel by construction.

Price band: $9.99 one-time for 5 Macs (DaisyDisk)20 to $65.99/yr (CleanMyMac Plus).21 Conversion mechanism is major-version exclusion on perpetual licences, stated in the vendor's own docs.23

Limit, and it is unresolved. Direct-store pricing could not be read: macpaw.com sits behind a challenge that returned HTTP 403 to a fetch and an unresolvable JavaScript challenge to a browser engine. Aggregators report $119.95 direct against the App Store's $89.99 one-time.21 That roughly $30 gap between channels is not reconciled here, and multi-Mac direct pricing is unverified.