macmon

macmon – Mac Monitor

macmon is a sudoless performance monitor for Apple Silicon Macs. It reads real-time CPU / GPU / ANE power usage, temperatures, and memory stats through a private macOS API — the same data powermetrics exposes — without requiring root access.

🌟 Features

• 🚫 Runs without sudo
• ⚡ Real-time CPU / GPU / ANE power usage
• 📊 CPU frequency-scaled and active ratios per cluster
• 💾 RAM / Swap usage and memory pressure
• 📈 Historical charts with average and max values
• 🌡️ Average CPU / GPU temperature
• 📋 Process list with per-process CPU, memory, power and GPU usage
• 🎨 Follows your terminal's color scheme
• 🪟 Can be displayed in a small window
• 🦀 Written in Rust

📥 Installation

Install macmon using brew:

brew install macmon

Install using MacPorts:

sudo port install macmon

Install using Cargo:

cargo install macmon

Install using Nix:

nix-env -i macmon

Tested on macOS 14 and later; older versions may work too.

🚀 Usage

Usage: macmon [OPTIONS] [COMMAND]

Commands:
  pipe    Output metrics in JSON format
  serve   Serve metrics over HTTP
  debug   Print debug information
  stress  Generate load for testing metrics
  help    Print this message or the help of the given subcommand(s)

Options:
  -i, --interval <INTERVAL>  Update interval in milliseconds [default: 1000]
  -h, --help                 Print help
  -V, --version              Print version

Interactive mode

Run macmon without a subcommand to open the terminal UI: the metric boxes on top and the process list below. Press ? in the app to see the keys.

Press / to filter the process list: commas separate alternatives, spaces combine words, ! negates a word. For example chrome, !helper, node mem:500m !user:root or app:safari, which shows the Safari app with its helpers. A plain word looks in the name or pid. Keys take a number, above it or below it with <: cpu:, gpu:, power:, mem: (k, m, g) and time: (s, m, h, d); or text: user:, path:, dir: (the working directory), cmd: (the command line, of your own processes) and app: (the name of the .app the process or its responsible process runs from); or an exact local TCP or UDP port, of your own processes: port:3000. Enter keeps the filter, Esc clears it.

Press Space to pause the process list: rows keep their order while values update, new processes wait until you resume, and the files list of i stops updating. Press f to turn the filter off and on without losing its text.

JSON output

You can use the pipe subcommand to output metrics in JSON format, which makes it suitable for piping into other tools or scripts. For example:

macmon pipe | jq

This command runs macmon in "pipe" mode and sends the output to jq for pretty-printing.

You can also specify the number of samples to collect using the -s or --samples parameter (default: 0, which runs indefinitely), and set the update interval in milliseconds using the -i or --interval parameter (default: 1000 ms). For example:

macmon pipe -s 10 -i 500 | jq

This will collect 10 samples with an update interval of 500 milliseconds.

HTTP server

You can use the serve subcommand to expose metrics over HTTP. This is useful for integrating with monitoring systems like Prometheus and Grafana.

macmon serve                   # default port 9090, interval 1000ms
macmon serve --host 127.0.0.1  # listen on localhost only
macmon serve -p 8080           # custom port
macmon serve -i 500            # sampling interval 500ms
macmon serve &                 # run in background

Two endpoints are available:

Endpoint	Format	Description
GET /json	JSON	Current metrics snapshot (same format as pipe --soc-info)
GET /metrics	Prometheus	Metrics in Prometheus text format

launchd service

To start macmon serve at boot, with nobody logged in, and keep it running:

sudo macmon serve --install                   # install and start (default port 9090)
sudo macmon serve --port 8080 --install       # with custom port
sudo macmon serve --host 127.0.0.1 --install  # listen on localhost only
sudo macmon serve --uninstall                 # stop and remove

--install needs sudo: it creates a LaunchDaemon at /Library/LaunchDaemons/com.macmon.plist that restarts on crash. The daemon runs as the user who called sudo, not as root, and with FileVault on it starts after the disk is unlocked following a restart. If that user has the agent below, the daemon replaces it.

To start at login instead, without sudo:

macmon serve --install --agent  # install and start (takes --host and --port too)
macmon serve --uninstall        # stop and remove

This creates a LaunchAgent at ~/Library/LaunchAgents/com.macmon.plist. Only one macmon service at a time: it's refused while the daemon is installed.

Prometheus and Grafana

The /metrics endpoint exposes metrics in Prometheus format. See examples/grafana for a local demo stack with Prometheus and Grafana.

# HELP macmon_cpu_temp_celsius Average CPU temperature in Celsius
# TYPE macmon_cpu_temp_celsius gauge
macmon_cpu_temp_celsius{chip="Apple M3 Pro"} 47.3

# HELP macmon_cpu_power_watts CPU power consumption in Watts
# TYPE macmon_cpu_power_watts gauge
macmon_cpu_power_watts{chip="Apple M3 Pro"} 8.42

# HELP macmon_fan_speed_rpm Fan speed in revolutions per minute
# TYPE macmon_fan_speed_rpm gauge
macmon_fan_speed_rpm{chip="Apple M3 Pro",fan="fan0"} 1234

# HELP macmon_cpu_scaled_ratio Combined frequency-scaled CPU ratio (0–1), weighted by core count
# TYPE macmon_cpu_scaled_ratio gauge
macmon_cpu_scaled_ratio{chip="Apple M3 Pro"} 0.037

# HELP macmon_cpu_active_ratio Combined CPU active residency ratio (not frequency-scaled, 0–1), weighted by core count
# TYPE macmon_cpu_active_ratio gauge
macmon_cpu_active_ratio{chip="Apple M3 Pro"} 0.092

# HELP macmon_cpu_tier_freq_mhz CPU tier frequency in MHz
# TYPE macmon_cpu_tier_freq_mhz gauge
macmon_cpu_tier_freq_mhz{chip="Apple M3 Pro",tier="E"} 1100
macmon_cpu_tier_freq_mhz{chip="Apple M3 Pro",tier="P"} 1800

# HELP macmon_cpu_tier_scaled_ratio CPU tier frequency-scaled ratio (0–1)
# TYPE macmon_cpu_tier_scaled_ratio gauge
macmon_cpu_tier_scaled_ratio{chip="Apple M3 Pro",tier="E"} 0.083
macmon_cpu_tier_scaled_ratio{chip="Apple M3 Pro",tier="P"} 0.015

# HELP macmon_cpu_tier_active_ratio CPU tier active residency ratio (not frequency-scaled, 0–1)
# TYPE macmon_cpu_tier_active_ratio gauge
macmon_cpu_tier_active_ratio{chip="Apple M3 Pro",tier="E"} 0.18
macmon_cpu_tier_active_ratio{chip="Apple M3 Pro",tier="P"} 0.04

macmon_ecpu_* and macmon_pcpu_* are still exported for the lowest and the highest tier. In a macOS VM the power and frequency gauges are left out.

Stress testing

Use macmon stress to generate load while checking metric behavior:

macmon stress
macmon stress pulse --duration 30
macmon stress cpu --duration 30
macmon stress cpu --workers 8 --duration 30
macmon stress gpu --duration 30
macmon stress all --duration 30

The default pulse mode generates a predictable cyclic CPU load with a fixed 50% duty cycle on half of the logical CPUs. The cpu and gpu modes continuously load only the selected processor, while all loads both. The cpu and all modes use all logical CPUs unless --workers is specified; --workers has no effect in gpu mode.

📊 Metrics

Power sources and macOS 27

macOS 27 broke the CPU/ANE power source that macmon previously used without root. The IOReport Energy Model counters can remain visible but stop updating: refreshing them now requires the private com.apple.private.pmgr.nrg.reporting entitlement. Apple's signed powermetrics has this entitlement and can keep using the named channels. Running macmon with sudo does not grant the same access.

The working rootless workaround we found is to read hidden energy reports from the AppleCLPC driver through IOReport. These reports use opaque IDs instead of discoverable CPU/GPU/ANE channel names, so macmon has to identify the counters and verify their units. This leaves macmon maintaining undocumented mappings for data that Apple's own tool can still read directly.

Both the CLI and the Rust library use known CLPC counters automatically, and Sampler::new() retains legacy fallbacks for older systems. New chips or driver versions may need additional mappings. On an unsupported configuration, zero CPU/ANE readings can mean unavailable counters rather than zero consumption. See docs/clpc-discovery.md for details.

Output format

The pipe command and the HTTP /json endpoint return the same metrics. Network, disk, uptime and battery are data only: there, in Prometheus and in macmon debug, not in the TUI. In a macOS VM ("vm": true in the soc object of pipe --soc-info and /json) power and frequencies read 0, temperatures null and fans are empty:

{
  "timestamp": "2025-02-24T20:38:15.427569+00:00",
  "temp": {
    "cpu_temp_avg": 43.73614, // Celsius, null when no sensor has a valid reading
    "gpu_temp_avg": 36.95167, // Celsius, null when no sensor has a valid reading
  },
  "memory": {
    "ram_total": 25769803776, // Bytes
    "ram_usage": 20985479168, // Bytes
    "swap_total": 4294967296, // Bytes
    "swap_usage": 2602434560, // Bytes
    "pressure_level": 0, // Kernel memory pressure level: 0 normal, 1 warning, 2 critical, null when unavailable
    "pressure_percent": 25, // Memory pressure as in Activity Monitor (0–100), null when unavailable
  },
  "fans": [
    { "name": "fan0", "rpm": 999, "max_rpm": 4900 },
    { "name": "fan1", "rpm": 1200, "max_rpm": 5200 },
  ],
  "cpu_scaled_ratio": 0.036854, // Combined frequency-scaled CPU ratio (weighted by core count, 0–1)
  "cpu_active_ratio": 0.092, // Combined active residency ratio (not frequency-scaled, weighted by core count, 0–1)
  "cpu_tiers": [ // CPU core types from the lowest to the highest: E, P on M1–M4; E, S on M5; P, S on M5 Pro/Max; E, P, S on M6; _ when the core type is unknown, e.g. the single tier of a macOS VM
    {
      "label": "E",
      "freq_mhz": 1100, // Tier frequency
      "scaled_ratio": 0.082656614, // Frequency-scaled ratio (0–1)
      "active_ratio": 0.18, // Active residency (not frequency-scaled, 0–1)
      "cores": [
        { "die_id": 0, "core_id": 0, "freq_mhz": 1600, "scaled_ratio": 0.14, "active_ratio": 0.24 },
        { "die_id": 0, "core_id": 1, "freq_mhz": 1700, "scaled_ratio": 0.12, "active_ratio": 0.2 },
      ],
    },
    {
      "label": "P",
      "freq_mhz": 1800,
      "scaled_ratio": 0.015181795,
      "active_ratio": 0.04,
      "cores": [
        { "die_id": 0, "core_id": 0, "freq_mhz": 2100, "scaled_ratio": 0.05, "active_ratio": 0.08 },
        { "die_id": 0, "core_id": 1, "freq_mhz": 2200, "scaled_ratio": 0.07, "active_ratio": 0.06 },
      ],
    },
  ],
  // Deprecated, use cpu_tiers: ecpu_* is the lowest tier, pcpu_* the highest
  "ecpu_freq_mhz": 1100,
  "ecpu_scaled_ratio": 0.082656614,
  "ecpu_active_ratio": 0.18,
  "pcpu_freq_mhz": 1800,
  "pcpu_scaled_ratio": 0.015181795,
  "pcpu_active_ratio": 0.04,
  "ecpu_cores": [/* cores of the first tier */],
  "pcpu_cores": [/* cores of the last tier */],
  "gpu_freq_mhz": 461, // GPU frequency
  "gpu_scaled_ratio": 0.021497859, // Frequency-scaled ratio (0–1)
  "gpu_active_ratio": 0.09, // GPU active residency ratio (not frequency-scaled, 0–1)
  "cpu_power": 0.20486385, // Watts
  "gpu_power": 0.017451683, // Watts
  "ane_power": 0.0, // Watts
  "all_power": 0.22231553, // Watts
  "sys_power": 5.876533, // Watts
  "ram_power": 0.11635789, // Watts
  "gpu_ram_power": 0.0009615385, // GPU SRAM power, Watts
  "net_in": 148210, // Bytes per second received by the Ethernet and Wi-Fi interfaces, null without one
  "net_out": 20544, // Bytes per second sent by the Ethernet and Wi-Fi interfaces, null without one
  "disk_read": 1048576, // Bytes per second read from the disks, without disk images, null when unavailable
  "disk_write": 262144, // Bytes per second written to the disks, without disk images, null when unavailable
  "uptime": 273600, // Seconds since boot, null when unavailable
  "battery_percent": 87, // Battery charge (0–100), null without a battery
  "battery_charging": true, // null without a battery
  "battery_plugged": true, // External power connected, charging or not, null without a battery
}

Active and scaled ratios

active_ratio is the share of the sampling interval during which the processor was doing any work. scaled_ratio is the same measure adjusted for operating frequency, showing the share of the processor's maximum possible capacity that was used.

For interval $T$, $t_i$ is the time spent doing work at frequency $f_i$, and $f_\text{max}$ is the hardware maximum frequency for the CPU cluster or GPU. The product $T f_\text{max}$ is the theoretical maximum frequency-time the processor could deliver at full load without thermal or power throttling:

R_\text{active} = \frac{\sum_i t_i}{T}
\qquad
R_\text{scaled} = \frac{\sum_i t_i f_i}{T f_\text{max}}

CPU ratios are calculated per core and averaged across the cluster. A core that did no work contributes zero; combined CPU ratios average all cores from all clusters:

R_\text{cluster} = \frac{1}{N_\text{cores}}\sum_j R_j

Examples:

• Full interval at half the maximum frequency → active = 1.0, scaled = 0.5.
• Half the interval at the maximum frequency → active = 0.5, scaled = 0.5.
• One of ten cores working at maximum frequency → cluster active = scaled = 0.1.

📚 Library usage

macmon can be used as a Rust library to collect Apple Silicon metrics in your own applications.

Add it to your project:

cargo add macmon --no-default-features

The default app feature enables the macmon executable and its terminal UI dependencies. Disable default features when using macmon only as a library.

Run the demo example from this repository:

cargo run --example demo

Then use the Sampler to collect metrics:

use macmon::Sampler;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut sampler = Sampler::new()?;

    // collect metrics over a 1000ms window
    let metrics = sampler.get_metrics(1000)?;

    println!("CPU power:  {:.2} W", metrics.cpu_power);
    println!("GPU power:  {:.2} W", metrics.gpu_power);
    match metrics.temp.cpu_temp_avg {
        Some(value) => println!("CPU temp:   {value:.1} °C"),
        None => println!("CPU temp:   N/A"),
    }
    println!("RAM usage:  {} / {} bytes", metrics.memory.ram_usage, metrics.memory.ram_total);
    for tier in &metrics.cpu_tiers {
        println!("{}-CPU:      {} MHz  {:.1}%", tier.label, tier.freq_mhz, tier.scaled_ratio * 100.0);
    }

    Ok(())
}

get_metrics(duration_ms) blocks the calling thread while collecting one IOReport delta over the complete interval. For a UI, server, or async application, create the sampler inside a dedicated worker thread and send the completed metrics back through a channel:

use std::{sync::mpsc, thread};

use macmon::Sampler;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let (tx, rx) = mpsc::channel();

    thread::spawn(move || {
        let mut sampler = Sampler::new().expect("failed to create sampler");

        while let Ok(metrics) = sampler.get_metrics(1000) {
            if tx.send(metrics).is_err() {
                break;
            }
        }
    });

    // Use recv() in a consumer thread or try_recv() in a non-blocking event loop.
    let metrics = rx.recv()?;
    println!("CPU power: {:.2} W", metrics.cpu_power);

    Ok(())
}

Creating Sampler inside the worker keeps its low-level macOS handles on that thread. In an async runtime, use its blocking-thread facility rather than calling get_metrics directly from an executor worker.

⚖️ Compared to htop and btop

htop and btop are good general monitors, and htop is lighter. On Apple Silicon, without sudo, macmon shows more:

• Real CPU and memory of other users' processes, such as WindowServer and root services. The kernel refuses them to htop and btop, which show 0; macmon reads them through the system's ps.
• CPU, ANE and total power, frequencies of each CPU cluster, and memory pressure.
• Power and GPU use per process.
• A process filter with queries (mem:>1g, app:safari), open files and sockets of a process.

Compared with htop 3.5.3 and btop 1.4.8 on macOS 27.

🔍 See also

• tlkh/asitop – The original tool. Written in Python, requires sudo.
• dehydratedpotato/socpowerbud – Written in Objective-C, sudoless, no TUI.
• op06072/NeoAsitop – Written in Swift, sudoless.
• graelo/pumas – Written in Rust, requires sudo.
• context-labs/mactop – Written in Go, requires sudo.

🤝 Contributing

Pull requests are welcome, including ones made with agents: point yours at agents.md.

When reporting an issue, mention the version from macmon -V, and for a wrong or missing metric add the output of macmon debug.

📝 License

Distributed under the MIT License.