Server setup

trailSailServer is a one-click installer for the Mac that runs the models. Double-click, one password, done: it checks Tailscale and Homebrew, installs Splash and the llama.cpp router, adds SearXNG for web search when Docker is running, generates the API keys, keeps everything running across reboots, and publishes it on your Tailscale network only. It finishes with a six-word pairing code.

This page is what the installer does, where it writes, and how to run it from a terminal.

What it installs

  • Splash — native MLX, one resident model. Apple M3 or newer, macOS 26.4 or later.
  • The llama.cpp router — GGUF models loaded on demand from the app.
  • SearXNG — in Docker, for the web search tool.
  • The control API — the small service the app talks to.

Each is published on your tailnet only, by tailscale serve:

EngineLocal portOn the tailnet
Splash8000https://<server>
llama.cpp router8081https://<server>:8443
SearXNG8888https://<server>:8444
Control API8086https://<server>:8446

where <server> is the machine's tailnet name (or its ts.net host).

The installer

Before anything runs it refuses with a plain message if macOS is older than 14, Tailscale is missing or not signed in, or Homebrew is missing. Splash needs an Apple M3 or newer and macOS 26.4 or later: on an older Mac the installer says so, leaves Splash out and installs the llama.cpp router alone. Your Tailscale account also needs HTTPS turned on, or the engines cannot be published: see Set up Tailscale first. It then:

  • sets the memory for models to suit the Mac's memory and installs the helper that lets you change it later — the one step that needs your password,
  • runs the install as you,
  • includes SearXNG when Docker is running, and
  • finishes with a dialog showing the pairing code.

The install log is /var/log/trailSailServer-install.log; the files live under /usr/local/trailSailServer.

Memory for models

Memory for models is how much of the Mac's memory a model may take on the GPU (macOS calls it the GPU wired memory limit). The installer sets the recommended amount: the Mac's memory less a reserve for macOS, which is 4 GB on a 16 GB Mac, 5 GB on 32 GB and 6 GB on 48 GB and up. A 32 GB Mac gets 27 GB, enough for a 27B model; a 64 GB Mac gets 58 GB. It is never lower than what macOS allows by default, and on a Mac with less than 16 GB it is left to macOS.

You can change it afterwards from the Server panel, from the menu bar item on the server, or with trailSailServer memory set:

  • If the Mac that runs the models is also the Mac you work on, choose less. The server suggests an amount for that: 20 GB on a 32 GB Mac, 8 GB on a 16 GB Mac.
  • It warns when the amount is too low for the model you use (the model will not fit, or will run slowly) and when it is too high for the Mac (macOS is left too little and the Mac can stop responding under load).
  • It cannot be set to within 2 GB of the Mac's memory.
  • A change counts for models loaded from then on. If you lower it while a model is loaded, restart that engine.

Setting it needs root, so the installer puts a small helper in place that does only this: it applies the amount at start-up and when you change it, and checks every amount itself. That is why the installer has to be run once on the server itself. A server that has only ever been updated from the app says "Run the trailSailServer installer once on this Mac" until then.

From the terminal instead

Unpack the tarball and run ./install.sh, with:

  • --with splash,router to skip SearXNG,
  • --model <hf repo> to download a specific model,
  • --language-only and --disable-ane for Splash memory controls,
  • --yes to skip the prompts.

Or install the package directly: sudo installer -pkg trailSailServer-<version>.pkg -target /.

What it writes

  • ~/.config/trailSailServer/config.json — host, engines, Splash model and flags.
  • Keys under ~/.config/trailSailServer/keys/ — mode 600; an existing Splash key is adopted so nothing breaks. Splash is started with its key, so it never serves without one.
  • launchd agents co.uk.yard8.splash, co.uk.yard8.llama-router and co.uk.yard8.control, so everything comes back after a reboot.
  • Logs in ~/Library/Logs/trailSailServer/.
  • The CLI at ~/.local/bin/trailSailServer.
  • The memory helper, as root: /Library/PrivilegedHelperTools/co.uk.yard8.trailsail-memory, its LaunchDaemon in /Library/LaunchDaemons, and the amount in /Library/Application Support/trailSailServer/memory-limit.

Re-running

Safe at any time. A rerun leaves running engines alone when nothing changed, and prints a fresh pairing code you can ignore if you are already paired.

Tailscale SSH (optional)

With sudo tailscale set --ssh on the server, the app's Server panel can push the package and later updates itself. The Mac App Store build of Tailscale cannot run Tailscale SSH; the standalone build from tailscale.com can. Without it, copy the package over by AirDrop, iCloud Drive or tailscale file cp.

Memory on 32 GB

Splash's 27B uses about 20 GB when busy; the router's 27B about 18 GB loaded. Keep one loaded at a time — the CLI warns when both are running.

Uninstall

/usr/local/trailSailServer/uninstall.sh removes the agents, the control API's serve mapping, the CLI and the memory helper, and puts the memory for models back to the macOS default. --purge also removes keys, pairings, the brew packages and the SearXNG container.

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