watt

A terminal system monitor that actually knows your GPU exists.

⚡ watt │ macbook │ up 4h32m │ cpu 12% │ mem 61% (19.2G/32.0G) │ 12 cores
╭─────────────────────────────────────────────────────╮
│ CPU 12.3%  ▁▂▁▃▂▁▂▅▃▂▁▁▂▃▅▇█▅▃▂▁▁▂▁▁▂▁▁▂▁▁▂▁▁▁▁ │
│ MEM 61.0%  ▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ │
│ GPU 24.1%  ▁▁▁▂▃▅▃▂▁▁▃▅▇▅▃▂▁▁▁▂▃▂▁▁▁▁▁▁▁▁▁▁▁▂▃▂ │
│ NET ↓1.2M  ▁▁▁▁▁▁▁▁▁▃▅█▅▃▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ │
│ PWR 8.2W   ▂▂▂▃▃▂▂▃▅▃▂▂▂▂▃▅▇▅▃▂▂▂▃▂▂▃▂▂▃▂▂▂▂▂▂▂ │
╰─────────────────────────────────────────────────────╯

other monitors just pretend GPUs don't exist?

Real-time CPU, memory, GPU, network, disk, and power monitoring with per-core heatmaps, process grouping, and Claude-powered system analysis. Catppuccin Mocha theme. On Apple Silicon, reads GPU stats directly from the hardware — no sudo, no kernel extensions.

cargo install --path .

NVIDIA GPU support:

cargo install --path . --features nvidia

watt                              # launch
watt -i 500                       # 500ms refresh interval
ANTHROPIC_API_KEY=... watt -p     # auto-analyze on startup

Chart — sparkline per metric, auto-scaled. Select with j/k, drill down with Enter. Mouse hover highlights, click drills down, scroll wheel navigates.

Processes — grouped by name with summed metrics. 47 Chrome Helper PIDs become one row showing Chrome Helper (×47). Sortable columns, filterable with /.

Profile — press P from any view. Sends an aggregated snapshot to Claude (requires ANTHROPIC_API_KEY) and displays a plain-language analysis of what's running and why. Scrollable with j/k.

no sudo? no kernel extension? just vibes?

macOS exposes GPU frequency residency counters through a private IOReport API — the same one Activity Monitor uses internally. watt subscribes to these counters and derives utilization from active-vs-idle time. No kernel extension, no entitlement, no sudo. Apple just left the door open.

• Utilization from frequency residency states (active vs idle time)
• Power from the Energy Model channel (millijoules → watts)
• Temperature from SMC sensor keys (Tg* prefix)

NVML via the nvidia feature flag. Per-process GPU memory shown in the process table.

forty-seven Chrome Helpers???

Processes are grouped by name into ProcessGroup structs via a HashMap. Each group sums CPU%, memory, and GPU memory across all PIDs sharing that name. The table shows count, aggregated metrics, and a ×N suffix.

it asks an AI what your computer is doing?

P or --profile builds a text prompt from the current snapshot — system stats plus the top 50 aggregated processes — and POSTs to the Claude Messages API (claude-sonnet-4-20250514, 1024 tokens). The response streams into a scrollable view. Requires ANTHROPIC_API_KEY env var.

src/
├── main.rs                 tokio select loop, channel wiring
├── config.rs               CLI flags (clap)
├── profile.rs              Claude API client, prompt builder
├── collect/
│   ├── mod.rs              snapshot collector (OS thread)
│   ├── process.rs          ProcessSnapshot, ProcessGroup, aggregate
│   ├── cpu.rs              per-core CPU
│   ├── memory.rs           RAM/swap
│   ├── gpu.rs              GpuBackend trait, NVIDIA impl
│   ├── apple_gpu.rs        Apple Silicon IOReport backend
│   ├── network.rs          rx/tx bytes
│   ├── disk.rs             disk I/O
│   └── power.rs            system power draw
└── ui/
    ├── mod.rs              terminal setup, draw dispatch
    ├── app.rs              state machine, keybinds, mouse
    ├── cpu_view.rs         sparklines, gauges, core heatmap
    ├── process_view.rs     grouped process table
    ├── profile_view.rs     Claude analysis view
    ├── header.rs           status bar
    ├── sparkline.rs        ring buffer
    └── theme.rs            Catppuccin Mocha palette

MIT