The Server panel

The Server panel (Settings → Server) polls the control API every ten seconds and shows the machine as it is — then lets you run it. From a Mac or a phone on the tailnet, it is your console for the machine that runs the models.

What it shows

  • package version, key version, macOS version, uptime, disk free, and the server's Tailscale state,
  • a memory bar — used, with wired memory marked, against the total and the memory for models.

Memory for models

How much of the server's memory models may use. The panel shows the amount in effect and lets you change it:

  • Recommended follows the server's own figure for its memory, for a Mac that only serves models. If the server is also the Mac you work on, it suggests a lower amount that leaves room for your apps.
  • The panel warns when the amount is too low for the model in use (it will not fit, or will run slowly) and when it is too high (macOS is left too little and the Mac can stop responding under load). All the figures come from the server; the highest is its memory less 2 GB.
  • A change counts for models loaded from then on. If you lower it while a model is loaded, the panel suggests restarting that engine.
  • If the server says the installer must be run once, do that on the server itself (the .pkg, or ./install.sh): changing this needs a small helper that only the installer can put in place. See Server setup.

Engine cards

One card each for Splash, the llama.cpp router and SearXNG: a status dot (running, starting, stopped, not installed), the model and flags, the address, and:

  • Start / Stop / Restart. Starting Splash waits for the model to load (up to four minutes) and shows the log lines as it goes.
  • Switch model… picks an installed model or a new repo. Splash gets Language only (less memory, no image input) and Disable the Neural Engine (steadier under memory pressure). Splash restarts with it; the next message uses the new model. For the router this sets the default only; loading stays in the model manager.
  • Install model… downloads without serving. Router: a temporary llama-server -hf … until the model is in the cache. Splash: a download subcommand if the installed Splash has one, else a second Splash on a spare port, with the main one stopped first so two models never share memory.
  • The installed models are listed with sizes; × deletes one from disk.
  • Logs shows the last 200 lines.

Paired devices

Lists every device with a token; × revokes one, which must pair again with a fresh code.

  • Pair another device… — mints a fresh code.
  • Rotate API keys… — new keys, engines restart, paired devices pick them up.
  • Update engines — brew upgrade and a SearXNG pull.
  • Install or update the package over SSH…
  • Forget this server…

A log pane under the panel shows what the server prints during any action.

Found a problem, or is something here unclear? Report it.