Moved: this repo was anphonic/llama-swatch until November 2026. Old links redirect here. Copies older than v0.4.4 can't find updates after the move, so download the latest release once by hand.

A desktop monitor for llama-swap. It shows what your llama-swap instance is doing: whether it's reachable, which models are loaded, which are busy or stuck, and a summary of throughput.

Works with llama-swap running anywhere you can reach over HTTP: the same machine, a box on your LAN, or behind a reverse proxy.

Tested with llama-swap v249 to v262. The header shows an amber note when the server reports a release outside that range.

Llama Swatch is an independent project. It is not affiliated with or endorsed by llama-swap or its maintainers.

_{Screenshots use made-up data.}

What the states mean

State	Looks like	Meaning
Not loaded	dimmed grey ring	Configured in llama-swap, not running
Loading	amber pulsing ring, seconds in the middle	llama-swap is starting the model. Past the slow-load warning (default 120 s) the card adds a slow load hint; llama-swap's own health-check timeout decides whether the load fails
Idle	solid green ring, TTL bar	Loaded, no requests in flight
Loaded	solid grey ring, no live data	Ready, but the monitor isn't reading live events, so it can't tell busy from idle
Busy	spinning blue arc, request count	At least one request is in flight and remains within its applicable timeout
Stalled	red ring with !	Unloading took too long, or the model stopped producing output
Unloading	fading grey ring	llama-swap is stopping the model

A busy card counts requests in three groups, plus how long ago output last arrived, e.g. 3 streaming · 1 started · 2 pending · last output 4 s ago (hover the card for the full wording):

• streaming: output is arriving.
• started: the reply has started (llama-swap has seen its headers) but no output has arrived yet. With llama.cpp this is usually brief; it shows up e.g. with other backends.
• pending: nothing has come back yet. A non-streaming chat ("stream": false) or an embeddings request looks like this until it finishes, because its whole reply arrives at once. With llama.cpp backends a streaming request also looks like this while it processes its prompt, since llama.cpp's server sends the reply headers only with its first result.

Stalls are judged for the whole model, not per request: while any request is streaming, the model is Stalled only when every streaming request has been quiet longer than the stream stall timeout, so a request waiting its turn never marks an active model as Stalled. With nothing streaming yet, it is Stalled once a started request has waited longer than the no-first-byte timeout without output. A pending request never makes the model Stalled, since the monitor cannot tell a long non-streaming generation or prompt from a stuck one; past the no-first-byte timeout the card adds a reply slow hint instead. So that setting drives the Stalled rule only for replies that have started, and otherwise drives the reply slow hint. A started request that waits for its first token more than 10 times that timeout while others stream adds a waiting long hint.

Busy and output-stall detection need llama-swap's live event stream (/api/events). The header badge shows its state:

• live: events are arriving and being read.
• waiting for events: the stream is open but no request event has been read from it yet (llama-swap may send nothing while idle). Ready models still show Idle unless an event fails to decode.
• events unreadable (amber): three request events in a row could not be decoded, so this llama-swap sends events in a shape this version of the monitor can't read.
• polling: no event stream.

Without readable events a ready model shows as Loaded rather than Idle. After events recover, a request that was already running may show as Idle until it sends output or finishes.

Toolbar and tiles

• Compact shows each model on one line (ring, name, state and the status text), so more fit without scrolling. Loaded only hides models that aren't loaded. Both are remembered.

• When nothing is loaded or loading, a Reload button offers the model that llama-swap's request history shows was used most recently, so you can bring it back with one click.

• Below the models, four tiles summarise traffic:

  • Requests: requests since llama-swap started.
  • Last input: prompt tokens of the newest request that reported token counts, with the cached tokens it reused underneath. Whether the input count includes the cached part depends on the backend (OpenAI-style usage counts the whole prompt; llama.cpp counts only the newly processed part), so the two are shown as reported and never added up.
  • Last output: output tokens of that same request, plus how much reply data has arrived so far while requests are streaming.
  • Generation p50: median generation speed, with p95 and p99, and a histogram underneath.

  Hover Last input or Last output, or tab to it, for the model behind that request and the totals since llama-swap started.

• In History, a token count llama-swap couldn't get from the backend shows as —.

Window

• On first launch the window opens centred at 481 x 770; on a screen whose work area is too short, the height shrinks to fit. It is always resizable, and your size, position and maximized state are remembered between launches. If the saved position is on a monitor that is no longer connected, the saved size is kept and the window opens at the default position on the main screen instead. After startup the app only resizes or moves the window when you switch to or from the Mini layout (see Skin below); leaving Mini puts back the full window's size, position and maximized state (or the default size and place if that spot is on a monitor that is no longer connected).

• Cards are one per row at a fixed height, and the header, toolbar, tiles and History columns keep their size as data changes, so nothing jumps around. Status banners and error toasts overlay the bottom of the window rather than pushing content down.

• The pin button in the header (left of the gear), or left of minimise on the Mini strip, keeps the window above other windows. Both buttons share one setting, saved in the settings file and re-applied at startup. Always-on-top is a window manager feature: on Linux it works on X11, but some Wayland compositors ignore it, in which case the button simply has no effect.

• When a newer release is out, a blue badge such as v0.2.0 available appears next to the connection status; click it to open the release page in your browser. The app never downloads or installs anything itself.

• Theme in settings (the gear) is Match system, Light or Dark. Match system follows the OS light/dark setting and switches live when it changes. On Linux it reads the desktop's dark-style switch through the XDG desktop portal (GNOME, KDE Plasma and most other desktops with a portal), falling back to the GTK theme name; if neither says, the app shows light. Pick Dark or Light to override it.

• Skin in settings picks the colours and layout: Standard (the default), High contrast, Muted (low-saturation), or Minimal (Standard colours in a compact strip layout). Your own skins are JSON files: drop *.json into the skins folder shown in settings (the Copy button copies its path), then press Rescan. Skins follow the Theme setting, so each has a light and a dark half. With the Minimal skin (or any skin with "layout": "mini") the window becomes a one-line strip; double-click or right-click it for Full dashboard or Settings (hover over the strip for a reminder). From the keyboard, tab to the strip and press the Menu key or Shift+F10; Escape closes the menu. Full dashboard returns to the full skin you last used. The "Cycle through loaded models in Mini" option rotates through every loaded model every 4 s.

  

  {
    "name": "Night Owl Mini",
    "layout": "mini",
    "light": { "bg": "#f6f7f9", "panel": "#ffffff", "text": "#1b1f24" },
    "dark":  { "bg": "#011627", "panel": "#0b2942", "text": "#d6deeb" }
  }

  name is required (at most 40 characters). layout is full or mini (default full). Colours are #rgb or #rrggbb only, from these 12 keys: bg, panel, text, muted, border, track, grey, green, amber, amberText, blue, red. Any key you leave out falls back to Standard, a missing dark block reuses light, and unknown keys are ignored. Files over 16 KB are skipped; skipped files are listed in settings with the reason.

Install

Releases are published on the Releases page, and only when every platform build succeeds. Pick the file for your OS:

OS	File
Windows	*-setup.exe (installer) or *.msi
macOS	*.dmg (universal: Apple silicon and Intel). Ignore *.app.tar.gz and any .sig files
Linux	*.deb, *.rpm, or *.AppImage

Each release also has a SHA256SUMS.txt. To verify a download, put it in the same folder as the file and compare:

# Windows (PowerShell): compare the output with the matching line in SHA256SUMS.txt
(Get-FileHash ".\<file>" -Algorithm SHA256).Hash.ToLower()

# Linux
sha256sum -c SHA256SUMS.txt --ignore-missing
# macOS
grep "<file>" SHA256SUMS.txt | shasum -a 256 -c

These builds are not code-signed or notarized, so each OS warns on first launch:

• Windows: SmartScreen says "Windows protected your PC". Click More info → Run anyway.
• macOS: right-click the app → Open → Open. If macOS says the app is damaged, run xattr -dr com.apple.quarantine "/Applications/Llama Swatch.app".
• Linux: use the .deb or .rpm, or chmod +x the .AppImage and run it. The API key is stored through the Secret Service (GNOME Keyring or KWallet), so one of those must be running.

Prefer not to trust a binary? Build it yourself, see Develop.

Connect

The API key is optional: you only need one if your llama-swap config sets apiKeys:.

1. Enter llama-swap's URL, e.g. http://llama-swap.example:8080. Paste the OpenAI-style .../v1 URL if that's what you have; the /v1 is removed.

2. If your llama-swap config has apiKeys:, paste one of those keys. It is stored in your OS keychain (Windows Credential Manager, macOS Keychain, Linux Secret Service), never in a file, and the UI never sees it again.

3. Save tests the connection first and only saves if it works.

   The saved key is only ever sent to the URL it was saved for. If you change the URL, paste the key again.

Check GitHub for a new release at startup is ticked by default in settings. Once per launch, after a connection has been saved and unless you've unticked it, the app makes one request to api.github.com for this repository's latest release tag. GitHub sees your IP address and the app version (in the User-Agent); nothing about your llama-swap is sent. Untick it to turn the check off. If you upgraded from a version without this setting, the check is on and the window says so once, with buttons to keep it on or turn it off.

The dashboard and Settings footers show Llama Swatch v…, the running app version used by the update check. The dashboard header's llama-swap version belongs to the server.

Thresholds (slow-load warning, no-first-byte timeout, and so on) are under Stall detection in settings. The no-first-byte timeout marks a model Stalled only when a reply has started but sent nothing; a request with no reply yet just gets the reply slow hint.

Develop

Requires Node 22+, Rust stable, and the Tauri prerequisites for your OS.

npm install
npm run tauri dev

cargo test --manifest-path src-tauri/Cargo.toml

Releases: push a v* tag. GitHub Actions builds Windows, macOS (universal) and Linux installers, publishes them as a release once all builds succeed (a tag containing a hyphen, like v1.0.0-rc1, is marked pre-release), and attaches SHA256SUMS.txt.

To build installers locally, run npm run tauri build.

Before a release, update the version in src-tauri/Cargo.toml, src-tauri/tauri.conf.json and package.json together (including their lockfiles). CI and release builds check that these three versions agree before installers can be published.

License

Licensed under the Apache License, Version 2.0. See LICENSE and NOTICE.

Third-party dependencies are under their own licenses (MIT, Apache-2.0, BSD, ISC, Zlib, MPL-2.0 and similar).