dashboard

A single-host computer dashboard and combined task manager: live system metrics, disk IO, process explorer, journal tail, and a systemd unit manager — all in one browser page. Built with FastAPI (Python, managed by uv) and htmx + Chart.js on the front end (no build step, JS is vendored).

Screenshots

Overview	Disks
Processes	Journal
Services	Plugins

Authorship

This code was written by Qwen3.8-27B (an LLM) and corrected by Johannes Schriewer hallo@dunkelstern.de.

If you want to see the OpenCode transcripts see the opencode sub-directory. I will include all transcripts when changing the software there.

⚠️ Security

This app has no authentication and no TLS. It binds to 127.0.0.1 by default and is meant to be used from the machine itself. To use it remotely, tunnel it — e.g. ssh -L 8501:127.0.0.1:8501 user@host and open http://127.0.0.1:8501 — or serve it through your own authenticated reverse proxy / VPN. Do not point DASH_HOST at a public interface without adding authentication yourself.

The Services tab can start/stop/enable/disable systemd units; any user who can reach the dashboard can do that (the app uses sudo for privileged actions, so the dashboard user needs sudo rights for systemctl).

Quick start

Requirements: Python ≥ 3.12 and uv. On Linux (it is tested on Arch).

uv sync
uv run python main.py

Then open http://127.0.0.1:8501.

Configuration

All settings are environment variables with the DASH_ prefix; a .env file in the project root is read automatically (see .env.example).

Variable	Default	Meaning
DASH_HOST	127.0.0.1	Bind address
DASH_PORT	8501	Port
DASH_SAMPLE_INTERVAL	2.0	Seconds between samples
DASH_RETENTION_MINUTES	60	In-memory history window (lost on restart)
DASH_LLAMA_BASE_URL	http://127.0.0.1:8080	llama-server (router mode) base URL
DASH_LLAMA_API_KEY	(empty)	Set if llama-server runs with --api-key
DASH_LLAMA_TIMEOUT	4.0	Seconds for llama-server requests

Tabs

• Overview — CPU + GPU (utilisation, CPU temperature, VRAM, GPU temperature when exposed by sysfs/hwmon), RAM + Swap + VRAM, active Wi-Fi connection (SSID via iw) and IP addresses of all up interfaces. History charts (CPU/GPU, memory/VRAM, disk I/O) are in-memory, sampled every DASH_SAMPLE_INTERVAL for DASH_RETENTION_MINUTES.
• Disks — partitions with size/use and per-device read/write rates.
• Processes — live table, filterable and sortable by CPU, memory, RSS, GPU and IO columns.
• Journal — streaming journalctl tail with level filter, unit filter, free-text search and a "hide sudo" toggle (on by default, since the dashboard's own sudo systemctl/journalctl calls otherwise log noise). Uses journal cursors, so filtering does not re-dump history.
• Services — all systemd service units with state and enabled-ness, sortable by name, state and enabled-ness; click a name for details (main PID, start time, restarts, recent journal lines) and run start / stop / restart / enable / disable actions.
• Plugins — currently llama.cpp: model status, load/unload buttons and a rescan for a llama-server running in router mode; sleep inhibitors: active block-mode systemd-inhibit locks with a verdict on whether the machine may sleep right now, plus a switch that makes the dashboard itself hold a sleep lock (released again on shutdown); and GPU power profiles (LACT): the active profile per GPU is polled, and each GPU gets a profile dropdown with a set and a reload button.

llama.cpp router mode

The plugin talks to a llama-server started with a models directory (router mode), which exposes the native /models, /models/load and /models/unload endpoints:

llama-server --models-dir /path/to/your/models --host 127.0.0.1 --port 8080

The plugin polls GET /health and GET /models, posts {"model": id} to /models/load and /models/unload, and rescans with GET /models?reload=1. If the server is down the plugin shows unreachable and the rest of the dashboard keeps working.

GPU power profiles (LACT)

The plugin shells out to lact cli (lact must be in PATH). It lists the GPUs with lact cli list, polls the active profile of every GPU with lact cli --gpu-id <id> profile get, and applies a selected profile with lact cli --gpu-id <id> profile set <name>. The per-GPU profile dropdowns are not refreshed automatically (LACT auto-switching can change the active profile behind the scenes, which the polled badge picks up); use the per-GPU reload button to refresh them.

Running as a systemd service

A ready-made unit is in deploy/dashboard.service:

You have to edit that file to point to the checkout of the tool and change the user and group!

cp deploy/dashboard.service /etc/systemd/system/dashboard.service
# adjust User= and paths if needed
uv sync   # once, after changing dependencies
systemctl daemon-reload
systemctl enable --now dashboard
journalctl -u dashboard -f

Project layout

main.py               # uvicorn entry point
app/
  config.py           # pydantic-settings (DASH_* env)
  main.py             # app factory, lifespan sampler
  sampling.py         # background sampler task
  state.py            # in-memory ring buffers
  collect/            # cpu / mem / gpu / disks / procs / net collectors (psutil + sysfs)
  systemd/units.py    # unit list / detail / whitelisted actions (sudo fallback)
  journal.py          # journalctl -o export parser + cursors
  render.py           # jinja env + filters
  routers/            # overview / disks / processes / journal / services / plugins
  plugins/            # base.Plugin + llamacpp + sleep + lact plugins
templates/            # htmx fragments
static/               # css, js, vendored htmx + chart.js
deploy/               # dashboard.service

Adding a plugin: create a module in app/plugins/ defining a Plugin instance (id, title, skeleton function) and register it in app/plugins/__init__.py. The skeleton is the static card shell, rendered once; it embeds the dynamically polled regions (e.g. a div with hx-get/hx-trigger="every Ns" pointing at the plugin's own endpoints).

License

The Unlicense — public domain dedication, no conditions.