lonoboard

A live dashboard for the machines running your AI models, drawn like a 1990s ANSI BBS.

Telnet in, or open a browser, and you get a live board for that machine. It shows the models it serves, their memory and GPU use, requests and tokens per second, an event log, and a chat window for talking to the models. Everything is drawn in CP437 and the CGA palette, the way ACiD and iCE drew it.

telnet gpu-box 2323           # any terminal; SyncTERM or NetRunner for the full effect
http://gpu-box:7777/          # the same board in a browser

It works with llama.cpp, Ollama, vLLM and any OpenAI-compatible server. It finds them on its own, and you can also list them in a config file. It runs on NVIDIA, AMD, Apple Silicon or a plain CPU. It's a single Python program with no dependencies.

Try it

pipx install lonoboard
lonoboard demo            # a simulated machine, with a fake chat
lonoboard check           # what lonoboard finds on this machine
lonoboard                 # the real board, on localhost

Or try it without installing: uvx lonoboard demo.

Python 3.11 or newer. Then open http://localhost:7777/ or run telnet localhost 2323.

The board

The board is made of screens, each sized to your window:

• Moving around:
  • A letter jumps to a screen. ← → flip between screens and esc goes back to main.
  • ↑ ↓ move the lightbar, enter opens the selected item, and 1–9 open a service.
  • A screen too tall for the window splits into pages. space reads on, b goes back.
  • The mouse works: click entries, rows and links, and use the wheel to turn pages.
• Details (enter on a service): status, model, context, last request, memory, the systemd unit or container, endpoints, launch flags, and a live tail of its log.
• Chat (c): talk to any chat model on the machine, and tab switches model. Replies stream, reasoning collapses into a summary, and esc stops a reply. Conversations aren't stored.
• Display (d):
  • IBM VGA 8x16 or 9x16 font.
  • VGA colour, amber or green monochrome.
  • iCE colours or blink for bright backgrounds.
  • CRT effect.
• ? opens help. q quits a telnet session after asking first.

Terminals

lonoboard negotiates echo, window size and terminal type, and adapts to what it finds:

Colour depth is chosen separately:

In a telnet session, press d and then m to cycle the colour mode, or u to switch between UTF-8 and CP437.

Model servers

lonoboard finds the systemd unit or Docker container behind each server. That gives it state, uptime, restarts and logs. See docs/sources.md for details.

Hardware:

• NVIDIA via nvidia-smi, including unified-memory systems like the DGX Spark.
• AMD via rocm-smi.
• Apple Silicon via ioreg.
• CPU-only.

Running it for real

lonoboard install --now          # a systemd user service running `lonoboard serve`

The board listens on localhost only until you say otherwise. To reach it from elsewhere:

• Tailscale (recommended). Keep it on localhost and publish it with tailscale serve. lonoboard then finds the other machines on your tailnet for the fleet screen.
• Your LAN. Set listen and a password. lonoboard won't listen beyond localhost or your tailnet without one. The password is asked for at a login prompt in the terminal, and the API takes it as a bearer token.

SECURITY.md covers what the board shows and how it's protected.

See docs/deployment.md for systemd, Tailscale, Docker and deploying to several machines over SSH, and docs/public-demo.md for hosting a public demo board.

Configuration

Nothing needs configuring, but everything can be. Settings go in ~/.config/lonoboard/config.toml, or in a file named by --config or $LONOBOARD_CONFIG:

[board]
tagline = "the attic box"

[[server]]
name = "lmstudio"
url = "http://127.0.0.1:1234"
kind = "openai"

examples/config.toml lists every setting, and docs/configuration.md explains them.

JSON API

Anything that doesn't ask for HTML gets JSON:

curl localhost:7777/              # host, services, other ports, fleet
curl localhost:7777/api/usage     # requests per hour, top 10

See docs/api.md for every endpoint.

Credits

• The fonts are PxPlus_IBM_VGA_8x16 and PxPlus_IBM_VGA_9x16, from The Ultimate Oldschool PC Font Pack by VileR. They are licensed CC BY-SA 4.0 and bundled, so the board works offline.
• The look is inspired by the ANSI art groups of the 1990s: ACiD, iCE, Blocktronics and the rest of the 16colo.rs archive.

License

MIT, see LICENSE. The bundled fonts keep their CC BY-SA 4.0 license.