The best thing in this macOS monitor is a chart that refuses to invent detail you did not record
A diagnostic performance logging utility for macOS: logs CPU, memory, GPU, network, and battery from the menu bar to help detect and diagnose performance issues. Free, open source, no telemetry.
At a glance
- What is it?
- A menu bar readout and background recorder that will not smooth over missing data, will not call modest memory growth a leak, and will not blur a short burst into a long-term average without showing you it did.
- Who is it for?
- Two things to know before you rely on it. The history only grows forward: readings recorded before a feature existed stay blank rather than being estimated, which means gaps in old data are permanent and you should not read them as zero.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly Swift, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- Zesty0wl/mac-performance-monitor1,017 ★ · +521
| Repository | 2026-09-17 | 2026-11-15 |
|---|---|---|
| Zesty0wl/mac-performance-monitor | 496 | 1k |
521 Stars in the last 59 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
It records, and recording is the feature
The framing in two lines: see what your Mac is doing now, then go back to the moment a slowdown or spike began. Then the list of what gets compared: processor, memory, graphics, network, disk, battery and sensor readings, against the processes behind them.
That is a different product from the one in the operating system. The built-in activity monitor is a live view. A live view is exactly what you do not want at the moment something goes wrong, because the interesting data was thirty seconds ago.
So there are three ways to use it: a live menu bar readout, a quiet background recorder, or a window you open when needed. The menu bar and the history recorder have independent switches, which is the detail that makes this usable on a laptop. Nobody wants the recorder running on battery and the menu bar item stealing space at the same time, and separating them means you can run the second for a week and drop the first.
The claim about data is short and specific: no usage telemetry, and recorded samples stay on your Mac. For a tool that holds a database of what your machine did every minute for weeks, that is the first thing to check rather than the last.
Charts that refuse to invent detail
The most valuable paragraph in the release notes is a short one about what the charts do not do, and I would put it on the wall of any monitoring tool.
Charts show an average over the recorded range, and the recorded range is drawn translucently behind it, so you can see how much time the average is summarising. Short bursts stay visible rather than being absorbed. Missing readings stay missing rather than being interpolated. Historical shapes remain stable as new data arrives rather than rescaling under you.
The one underneath those is the strongest: zooming does not invent detail. Older data keeps the resolution it was recorded at, so if you recorded one sample a minute in March and one a second last week, zooming into March will not manufacture a smooth curve from five points.
This is the difference between a tool you can reason from and one you cannot. A monitoring app that interpolates gaps will cheerfully draw you a beautiful ramp during the period your machine was asleep, and you will spend an afternoon debugging your application instead of your monitor.
The same honesty appears in the data model. New history grows from new recordings, and older logs stay as gaps for fields they never stored. New database fields preserve detail from new samples and cannot restore missing older readings. Hardware inventory is a current snapshot kept separate from history.
So the practical advice when reading its charts is the advice its author is giving: a blank region means not recorded, not zero.
The alert rules are the second thing worth reading closely
The alert design is a direct argument against how most monitoring tools behave, and the opening sentence is the thesis: high swap usage alone is not a reason to warn you.
Swap use is the metric every system monitor alerts on, and it is almost always a bad alert. Swap rises when the machine has been busy; it does not fall promptly afterwards, so a single busy afternoon leaves you permanently over a threshold and permanently notified.
The rules here look for continued growth or paging strain instead, and they can report further worsening without first waiting for usage to drop below an old fixed limit. That is a rule about the shape of a curve rather than its height.
The process checks are more careful still. Growth that has settled is rejected. Modest findings stay as quiet observations rather than being called a memory leak. A separate fast-growth check catches runaway use without waiting out a long window.
That last distinction, leak versus high usage, is the one that matters most. Almost every system tool that claims to have found a memory leak is reporting high usage. A genuine leak has a growth curve; high usage is a level.
Then the notification design, which is unusually well thought out. The menu bar badge shows active issues and their evidence. You can snooze an ordinary alert for an hour or open it in the workspace at the relevant moment. Related memory alerts share a single notice rather than arriving separately, and only critical ones request sound.
That is alert design for humans rather than for dashboards, and it is the kind of thing you only get right if you have been on the receiving end.
Two features that hand the diagnosis to something else
The newest release adds two ways to get an explanation rather than a chart, and both have unusually clear boundaries drawn around them.
The first asks a plain question, such as why the machine is slow, or lets you tap an area like memory or battery. The app checks its own readings and history, then a model running on the machine explains them in plain language, with links to the matching charts. Nothing leaves the machine. And the app states that answers can still be wrong, so the facts behind each one are one click away.
That is the right design for a feature like this: the explanation is a convenience layer over data you can inspect, not a replacement for it. And it works without the vendor's assistant layer enabled, which means the summaries are not silently dependent on a feature flag.
The second hands a harder question to a coding agent. The app copies a prompt explaining the recorded history, or gives you a one-line command to connect a read-only protocol server:
claude mcp add mac-performance-monitor -- "/Applications/Mac Performance Monitor.app/Contents/MacOS/mpm" mcpRead-only is the operative word, and it is the right constraint for a tool that hands its database to a language model. The agent can read your machine's performance history and cannot change it, which means the failure mode is a bad answer rather than a changed system.
The other new feature is narrower. Right-clicking a process shows when the recorder observed it running, with optional detail about operating-system app and media activity. That optional part needs broad disk access and a per-window opt-in, and the label says plainly that it is not verified foreground-use history.
Four separate hedges in one feature description. Whoever wrote this documentation does not want you to file a bug about something the app never claimed.
Battery features that admit they are estimates
The energy features occupy more release-note space than anything else, and they are worth reading as a case study in how to present uncertain hardware data.
Runtime comes from operating-system estimates or from recent steady use, and the forecast is drawn as a dashed line kept visually separate from real charge data. The stated caveat is the important part: it assumes the same workload, and it is not a fixed promise of how long the battery will last.
Battery health and cycle history show daily data over months and years, kept separate for each battery pack. That last detail is the one that catches people: a machine that has had its battery replaced has two histories, and merging them produces a nonsense curve.
The accessories panel reports what the operating system reports for mice, keyboards and earphones. Optional low-battery alerts check at most once a minute, and unknown values stay unknown rather than being shown as zero.
The graphics readings follow the same pattern. Total, reads and writes share a historical bandwidth chart that is marked as a preview because the values are estimates derived from coarse system buckets. Neural Engine time and power are separate readings, each with its own requirement: one needs a recent operating system version, the other needs an approved privileged helper.
Calling a chart a preview because the underlying source is coarse is the single most useful thing in these notes. It tells you the number is derived, so you can decide whether a trend is meaningful before you go looking for a cause.
A Swift package with the release engineering in plain sight
The repository is a Swift package, which means the build definition is a file in the repository rather than project settings hidden inside an application. For a project that publishes builds, that is the difference between reviewing the build and trusting it.
The file list shows a mature process rather than a hobby project. There is a package manifest and a resolved dependency file, which together mean your build pins the versions it was tested against. There is a directory of git hooks installed into your checkout. There is a configuration for the language formatter, which means formatting is enforced by a tool rather than by reviewer comments.
Then the release apparatus, and it is extensive: a changelog, a separate release notes document, a release checklist in the documentation, a packaging directory, a cask directory, scripts, and a checks directory. A cask directory plus a Homebrew cask badge means the install path is a package manager rather than a download-and-drag.
For a macOS app distributed outside the app store, a cask is the correct mechanism: it is declarative, it is reviewable, and it means your version number is machine-readable rather than something a human typed into a web page.
The localisation setup is the other sign of a project with real users. There is a localisation directory, a translation guide, a glossary, and a configuration file pointing at a translation platform. A glossary in particular tells you the project has fought the problem of the same word meaning different things in different languages, which is the problem that makes most software localisation quietly bad.
Then the governance files: a code of conduct, a contributing guide, a security document, and a per-project agent instruction file. The security document is the one to read before sharing a trace file, which the upgrade notes point you to specifically, and that is a sensible thing to point at: trace files contain a record of what your machine did, which is not something you want attached to a bug report by accident.
Editorial conclusion
Two things to know before you rely on it. The history only grows forward: readings recorded before a feature existed stay blank rather than being estimated, which means gaps in old data are permanent and you should not read them as zero. And several of the newest readings are explicitly labelled as estimates derived from coarse system buckets, with the Neural Engine items needing a recent operating system version and an approved helper. Both of those are the app telling you the truth about its own limits, which is rarer than it should be in a diagnostic tool and is the reason to trust the parts it does measure.
Frequently asked questions
What does the Mac Performance Monitor record?
Processor, memory, graphics, network, disk, battery and sensor readings, alongside the processes behind them. It runs as a live menu bar readout, a quiet background recorder, or a window you open when needed, and the menu bar and recorder have independent switches so you can leave one off.
Does it send any data anywhere?
No. The readme states there is no usage telemetry and recorded samples stay on your machine. It also bundles a read-only protocol server for coding agents rather than a read-write one, so an agent you point at it can read your performance history and cannot change it.
How does it handle missing historical data?
It does not fill it in. Missing readings stay missing, short bursts stay visible rather than being averaged away, and zooming does not invent detail, so older data keeps whatever resolution it was recorded at. New database fields preserve detail from new samples and cannot restore older readings, so a blank region means not recorded rather than zero.
Why does it not alert me about high swap usage?
Because swap use rises when a machine is busy and does not promptly fall afterwards, so a single busy afternoon leaves you permanently over a threshold and permanently notified. The rules instead look for continued growth or paging strain, reject growth that has settled, keep modest findings as quiet observations rather than a leak diagnosis, and have a separate fast-growth check for runaway use.
Which readings are estimates rather than measurements?
The graphics bandwidth history is marked as a preview because it is derived from coarse system buckets. Runtime uses operating-system estimates or recent steady use, and its forecast is drawn separately from real charge data with the caveat that it assumes the same workload. Neural Engine power needs an approved privileged helper, and Neural Engine time needs a recent operating system version.
How do I install it?
Through a Homebrew cask, which is the packaged route the repository ships, or from a release build. The project is a Swift package with a resolved dependency file and formatting enforced by a tool, and there is a separate release checklist and a security document you should read before sharing trace files, since those contain a record of what your machine was doing.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/zesty0wl-mac-performance-monitor)Star history chart for your README
[](https://hysenlabs.com/projects/zesty0wl-mac-performance-monitor)Was this useful?