Tempra

Your Mac's vital signs, live, over time, and a step ahead.

A native macOS menu bar app that monitors CPU, GPU, memory, temperatures, battery and power, keeps a long-term history, and forecasts what comes next with Mistral's Timestral, running entirely on your Mac.

Features · Screenshots · Install · Forecasting · Architecture · Roadmap

---

Features

🧭 Lives in the menu bar	An icon plus one live value of your choice (CPU, GPU, memory pressure, temperature, fan speed, battery or power). The icon follows the statistic you show.
🚦 Health at a glance	Every key metric is Normal, Warning or Critical, using macOS's own signals where they exist. Warning values turn orange, Critical red, in the panel, the dashboard and the menu bar.
🪟 Glanceable summary	Click the icon for a compact panel with the essentials, sparklines and the latest forecasts.
📊 Full dashboard	A real window with history from 15 minutes to 30 days, per-core CPU (P and E cores), GPU renderer and tiler load, memory breakdown, thermals, battery and power, disk and network, and a live process list.
🔮 Forecasts with Timestral	Multivariate, probabilistic forecasts (10th–90th percentile bands): time to empty or full and battery life if you unplug, next-hour CPU, GPU, memory pressure, temperature and power, and a 3-year battery health projection whose range narrows as history grows.
🗄️ Long-term history	Every sample for 24 h, 1-minute averages for 30 days, hourly averages forever, in a single SQLite file.
🪶 Lightweight	Native Swift, no daemon with root privileges. Charts only redraw while on screen, so Tempra uses about 1–2% of one core in the background; the forecaster a few milliseconds per minute.
🎨 Liquid Glass icon	Designed in Icon Composer: a history line flowing into three forecast bands, in six palettes (Indigo, Night, Dawn, Aurora, Sky, Graphite) to pick from in Settings.
🔒 Private	No accounts, no telemetry, no network calls except the one-time model download.

What Tempra measures

Area	Metrics	Source
CPU	Total, user and system load, per-core load (P/E), load average	host_processor_info
GPU	Utilization, renderer, tiler, unified memory in use	IOAccelerator performance statistics
Memory	Pressure, used, app, wired, compressed, cached, swap	host_statistics64, sysctl
Thermals	SoC average and hotspot, battery, SSD, thermal state	Apple Silicon HID temperature sensors
Fans	Speed of each fan, with its minimum and maximum	System Management Controller (SMC)
Battery	Level, charging state, time remaining, cycles, health, capacities, voltage, current	IOPowerSources, AppleSmartBattery
Power	Whole Mac, chip (SoC), CPU & rest of chip, GPU, adapter input — live, every sample	SMC (PSTR, PHPC, PDTR), IOReport GPU energy
Battery flow	Watts into or out of the battery, adapter rating	AppleSmartBattery
I/O	Network down/up, disk read/write, free space	getifaddrs, IOBlockStorageDriver
Processes	CPU over the last second and memory footprint per process, as in Activity Monitor	top

Health states

Metric	Source	Warning	Critical
Memory	Available memory (the pressure Activity Monitor plots) and swap activity	Pressure above 70% or swapping out over 1 MB/s, for a minute	Pressure above 85%, swapping over 20 MB/s, or macOS critical
Temperature	macOS thermal state	Serious (throttling)	Critical
Battery level	macOS low-battery warning	Early (below ~22%)	Final (~10 min left)
Battery health	macOS battery condition, Apple's 80% threshold	Capacity below 80%	Service recommended
CPU, GPU	1-minute average load (GPU: only while it draws 2 W or more)	75–90%	above 90%
Power	Runtime at the last minute's average draw, or adapter headroom	Under 2 h left, or above 90% of the adapter's rating	Under 1 h left, or battery draining while plugged in because the adapter can't cover the draw (not when macOS uses the battery on purpose)
Startup disk	Free space	Under 20 GB or 10%	Under 5 GB or 3%

Normal is left uncoloured so that Warning and Critical stand out; the state is always written out too. Whatever the menu bar shows, a red dot on its icon flags any other metric in Critical — for example memory while the menu bar shows CPU. A state that gets worse shows at once; one that improves waits two minutes, so a metric hovering around a threshold doesn't make the colours flicker.

Screenshots

Menu bar panel: the essentials at a glance

Dashboard overview: all vitals with history

CPU: history and per-core load	GPU: utilization, renderer and tiler
Battery & power: health, watts, time remaining	Memory: pressure, breakdown and swap

_{Forecasts: every Timestral forecast with its uncertainty band and what it was conditioned on}

Requirements

• A Mac with Apple Silicon (M1 or later) running macOS 14 Sonoma or later
• uv for the optional forecasting engine (brew install uv)
• To build from source: Xcode 26+ (for Icon Composer icons) and XcodeGen

Installation

Download

1. Download Tempra-x.y.z.zip from the latest release and unzip it.

2. Move Tempra.app to /Applications.

3. Open it. Tempra isn't notarized by Apple yet, so the first time macOS will refuse to open it: go to System Settings → Privacy & Security, scroll down and click Open Anyway. Alternatively, from Terminal:

   xattr -dr com.apple.quarantine /Applications/Tempra.app

Build from source

git clone https://github.com/frankydedo/Tempra.git && cd Tempra
scripts/build.sh --install     # builds Release, copies to /Applications and launches

scripts/package.sh produces the release zip and its SHA-256 checksum in build/.

First run

Open Settings → General to choose what appears in the menu bar and turn on Open Tempra at login.

To work on the code in Xcode:

xcodegen generate && open Tempra.xcodeproj

Forecasting with Timestral

Timestral is Mistral's family of zero-shot time-series foundation models. Tempra runs it locally in a small Python worker. It needs no training on your data: it reads your recent history and predicts the next steps, with uncertainty bands.

Enable it from Settings → Forecasting → Install, or from Terminal:

forecaster/setup.sh ~/Library/Application\ Support/Tempra/forecaster

This creates a private virtual environment with PyTorch and Timestral, on a native arm64 Python even when uv itself runs under Rosetta. The model weights (21 MB for the 5M model) download on first run.

How the forecasts work

Your Mac's vitals are tightly linked: load drives power and temperature, power and the adapter drive the battery. Tempra's forecasts use those links instead of predicting each metric on its own.

Block	Steps · context · horizon	Targets	Conditioned on
System	1 min · 14 days · 1 h	CPU, GPU, memory pressure, SoC temperature, fan speed and system power, forecast jointly as variates of one series	Adapter state (past covariate), time of day (known-future covariate)
Battery	5 min · 30 days · 12 h	System power, forecast jointly with CPU and GPU load	Power-source scenario and time of day (known-future covariates)
Battery health	1 week · all · 3 years	Maximum capacity	Capacity lost per charge cycle and cycles per day (see below)

Battery level comes from a hybrid model: Timestral forecasts the power the Mac will draw, and Tempra converts it into battery percentage using the battery's actual energy (mAh × V). This stays physically consistent from day one, even before your history contains enough discharges for a purely statistical model. Depending on the state, Tempra shows:

State	Shown as
On battery	"Battery until ~18:40 (17:50–19:30)"
Charging	"Full at ~11:25 (11:15–11:40)", plus until when it would last if you unplugged now
On adapter, not charging (e.g. held at the charge limit)	"Held at 80% · on battery until ~22:30"

While charging, the battery level follows the same idea: the power going into the battery is the smaller of what the battery accepts at its current level (constant current up to ~80%, then tapering) and what the adapter has left after the system's own consumption, which Timestral forecasts. The acceptance curve is learned from your own charging sessions, battery level by battery level; until there is enough history, a standard Li-ion taper scaled to the current charging power is used (calibrated to macOS's own estimate right after plugging in, when the battery gauge hasn't updated yet). The adapter's rating is read from the charger itself (USB-C Power Delivery or MagSafe).

The battery chart shades charging periods in green and time on the adapter without charging (for example held at 80% by optimized charging) in yellow.

Battery health over three years

Battery wear is slow (a few tenths of a percent a month) and long-term, so instead of a time-series model Tempra projects it from two rates:

capacity in t days = capacity now − capacity lost per cycle × cycles per day × t

Both start as wide estimates — capacity lost per cycle from the battery's lifetime average (capacity lost so far ÷ cycles so far), cycles per day from a vague prior — and are refined with this Mac's own history (the trend of capacity against cycle count, and the cycles actually used each day). The 10–90% range comes from sampling both rates, so it is wide on day one and narrows as days and cycles accumulate: in a simulation, from about 34 points of range after two days to 5 after six months and under 4 after a year. Tempra shows the expected capacity in three years and when it should reach 80%, the level below which Apple recommends service.

Do forecasts get better over time?

Yes, up to a point. Timestral is zero-shot: its weights never change and it is not trained on your data. What grows is the history it reads. With two weeks of context for the system block it can see daily and weekly habits; with 30 days of battery history, your real discharge patterns. Beyond those windows older data is not used, so forecasts keep adapting to how you use your Mac now.

Accuracy

Tempra checks its own forecasts. It keeps a copy of each forecast every 15 minutes and, once the horizon has passed, compares it with what actually happened. Forecasts → Accuracy shows, for the last 7 days:

• Typical error: mean absolute difference between forecast and reality, in the metric's unit.
• vs. "no change": how much better (or worse) Timestral did than simply assuming the current value stays put.
• Inside band: how often reality fell inside the 10–90% band. A well-calibrated forecast is close to 80%.

Bands calibrate themselves: every scored forecast widens or narrows its metric's band a little, so that over time about 80% of what actually happens falls inside it.

Battery forecasts are scored only while the adapter state stays the same as when they were made, and the "held at the charge limit" state is not scored, since it isn't a real prediction.

Forecasts appear after about an hour of history (two hours for battery; the battery health projection from the first day) and refresh every minute. On an M4 a full refresh takes under a second of CPU time. You can switch between Timestral 5M and 28M in Settings.

Architecture

flowchart LR
    subgraph App["Tempra.app (Swift)"]
        S[Sampler<br/>CPU · GPU · RAM · HID · IOKit] --> M[SystemMonitor]
        M --> UI[Menu bar · Summary · Dashboard]
        M --> DB[(history.sqlite)]
        F[ForecastService] --> UI
    end
    subgraph Worker["Forecaster (Python)"]
        W[worker.py<br/>Timestral] 
    end
    DB -- history --> W
    W -- forecasts --> DB
    DB -- forecasts --> F
    F -. launches .-> W

Loading

Tempra/
├── App/            App entry point, menu bar label, settings keys
├── Metrics/        Readers for CPU, GPU, memory, thermals, battery, I/O, processes
├── Storage/        SQLite history (raw → minute → hour) and metric definitions
├── Forecast/       Launches the Python worker and reads its forecasts
├── Resources/      AppIcon.icon (Icon Composer) and alternate icon renders
└── UI/             Summary panel, dashboard sections, charts, settings
forecaster/         Timestral worker, pyproject and setup script
design/             Icon generator, Icon Composer documents and previews
scripts/build.sh    Release build and install

Privacy

Everything stays on your Mac, in ~/Library/Application Support/Tempra/. Tempra makes no network requests. The forecaster downloads the model from Hugging Face once and then runs offline. You can delete all history from Settings → Data.

Roadmap

Metrics

• Live power by component without privileges: whole Mac, chip, CPU & rest of chip, GPU
• Neural Engine and memory power (macOS only exposes them to privileged processes)
• CPU cluster and GPU frequencies, GPU temperature
• Who's consuming: the processes behind CPU, GPU and power spikes, recorded during the spike and annotated on the charts, plus an energy ranking per app
• Intel Mac support (SMC thermals)

Forecasting

• Multivariate forecasts with covariates (adapter state, time of day) and a physics-based battery model
• Clock-time forecasts: "battery until ~18:40", "full at ~11:25", and "will it last until 19:00?" answered with a chance from the forecast bands
• Battery report: daily and weekly time on battery, real runtime per full charge, charging habits (e.g. hours spent at 100%) with battery-health tips
• Startup disk forecast: when the disk will be full at the current rate
• Run Timestral natively through Core ML on the Neural Engine, removing the Python dependency
• Forecast accuracy report comparing past forecasts with what actually happened

App

• Desktop and Notification Center widgets (WidgetKit): CPU, memory, battery with its forecast, temperature
• Export history to CSV and JSON
• Customisable dashboard layout and summary tiles
• Italian localization
• Signed and notarized releases, auto-updates (Sparkle) and a Homebrew cask
• Unit tests and GitHub Actions CI

Known limitations

• Temperatures come from private IOKit HID APIs. They work on current Apple Silicon Macs, but a macOS update could change sensor names. For the same reason Tempra is not sandboxed and cannot ship on the App Store.
• GPU temperature is not exposed separately on Apple Silicon, so it is included in the SoC readings.
• Process CPU is a percentage of one core and memory is the physical footprint, as in Activity Monitor.

Development notes

Tempra logs state changes (adapter plugged or unplugged, health state transitions with the values behind them, forecaster lifecycle, battery forecast changes) to ~/Library/Logs/Tempra/Tempra.log and to the macOS unified log. The forecaster writes to ~/Library/Logs/Tempra/forecaster.log.

tail -f ~/Library/Logs/Tempra/Tempra.log
/usr/bin/log stream --predicate 'subsystem == "com.fdedominicis.tempra"'

The app icons are Icon Composer documents generated from code, so every variant shares the same geometry:

python3 design/make_icons.py   # writes design/icons/*.icon and Tempra/Resources/AppIcon.icon (default)
design/export_icons.sh         # renders them with ictool for the Settings picker and docs/icon.png
swift design/contact_sheet.swift design/icon-variants.png   # side-by-side preview

Open any .icon in Icon Composer (bundled with Xcode) to fine-tune its Liquid Glass settings.

Debug builds can render every view to PNG without screen-recording permission, which is how the screenshots above were produced:

build/DerivedData.noindex/Build/Products/Debug/Tempra.app/Contents/MacOS/Tempra -TempraSnapshotDir docs/screenshots

License

Tempra is released under the MIT License.

Timestral's code is Apache 2.0 and its model weights are under Mistral's Modified MIT license. Neither is included in this repository: the forecasting engine downloads them when you install it.

Acknowledgements

• Mistral AI for Timestral
• Stats and asitop, which document much of Apple Silicon's sensor landscape