Splash Launcher

A native macOS app for starting, stopping and monitoring a local Splash model server, without the command line. Built and tested for Splash-M1, the community port of Inco's Splash for M1 and M2 Macs.

Unofficial. Splash Launcher is an independent project. It is not made by, affiliated with, or endorsed by the developers of Splash. It controls an existing Splash installation and never installs, updates or modifies Splash itself.

Install

1. Install Splash-M1 by following the instructions at paperniuk/splash. Splash Launcher controls an existing installation and doesn't install Splash for you; if it can't find splash-m1, it says so and links to that page.
2. Get Splash Launcher, either way:

   • Download the zip from the latest release, unzip it and move Splash Launcher to Applications. The app isn't notarised by Apple, so the first time you open it macOS blocks it: open System Settings → Privacy & Security and click Open Anyway. Each release is built by GitHub Actions from that version's source.

   • Or build it from source with Xcode 15.3 or later, or the Command Line Tools (xcode-select --install):

     git clone https://github.com/alphaexposure/splash-launcher.git
     cd splash-launcher
     zsh Scripts/build-app.zsh
     open "dist/Splash Launcher.app"

     To update, run git pull and zsh Scripts/build-app.zsh again.

How it fits together

• Splash is Inco AI's local inference engine for M3 and newer Macs.
• Splash-M1 is a community port of Splash for M1 and M2 Macs. You install it separately.
• Splash Launcher (this app) starts, stops and monitors Splash-M1. It doesn't run models itself, and your chat apps talk to the Splash-M1 server directly, not through the launcher.

Features

• Start, stop and restart Splash with one click, and open its web chat page.
• Settings: model, host, port, reasoning effort, KV cache format, memory and context limits, model aliases, and an optional API key stored in the macOS Keychain.
• Live overview:
  • Readiness, memory use and limits, and context limit, as reported by Splash.
  • Token activity for the whole server session.
  • A Latest request panel with prompt, cached and output tokens, time to first token, decode speed, and derived prefill, decode and total times.
• Installed models: each model's real size on disk (downloads plus the weights Splash prepared from them), its format, and a button that opens the folder holding its files. Splash's prepared-weights cache is shown too, including leftovers from removed models.
• Server output log that fits the window, with auto-scroll, Copy, download progress bars that update in place, and any API key redacted.
• Menu bar control to reopen the window or stop the server.
• Optional extras: launch at login, and a light, dark or system appearance.

Requirements

• macOS 14 or later on an M1 or M2 Mac.
• An existing Splash-M1 installation (paperniuk/splash). The launcher looks for splash-m1 in /opt/homebrew/bin and ~/.local/bin, or you can choose the executable in Settings.
• Official Splash (M3 or newer) uses the same API and may work if you choose its splash executable in Settings, but it hasn't been tested.
• To build: Xcode 15.3+ or a Swift 5.10+ toolchain.

Development

Scripts/build-app.zsh builds a release binary and packages it as dist/Splash Launcher.app, ad-hoc signed for local use. swift build and swift test compile and test the code but don't update the app bundle. Running the tests needs full Xcode, because the Command Line Tools don't include XCTest.

swift test

To publish a release, push a version tag such as v0.2.0. The Release workflow then runs the tests, builds the app with that version number, and attaches the zip and its SHA-256 checksum to a GitHub Release.

How it behaves

• Local by default: binds to 127.0.0.1 on port 8000, so only this Mac can connect. Change the host deliberately if you want other devices to reach it.
• Only stops what it started: the launcher stops a Splash process only if it launched that process itself. A server started elsewhere is never stopped.
• No shell: Splash is launched directly with an explicit argument list, never through a shell.
• Keys stay hidden: the API key is kept in the Keychain, passed to Splash through its environment, and redacted from the log view.
• Latest request panel: this reads the summary line Splash prints after each request. It only works for servers this app started. Values marked "≈" are calculated from that line, not measured directly.

Acknowledgements

Splash Launcher builds on Splash by Inco AI and the community Splash-M1 port for M1 and M2 Macs. Both are independent projects; Splash Launcher is not affiliated with either.

Credits

The app icon concept (an upright "S-Wave") was developed with design assistance from ChatGPT. The mark is drawn in code from a single source, SplashMark.swift, which produces the app icon, the sidebar mark and the menu bar symbol.

License

MIT