LlamaApp

A Windows 11 system-tray app for running llama.cpp models locally — a WinUI 3 port of the macOS llama.app menu-bar app.

The app lives in the notification area. Clicking the tray icon pops a borderless, Mica-backed flyout (anchored to the icon) showing your locally available models and a recommended catalog from the Hugging Face Hub. It adopts, launches, or auto-installs a llama binary and drives it over its local REST API.

Features

Tray-only flyout UI — no taskbar window; left-click the tray icon to show/hide, right-click for Open/Exit. Mirrors the macOS menu-bar experience on Windows 11.
Auto-managed llama.cpp binary — adopts a running server on localhost:2276, launches %USERPROFILE%\.llama-app\llama.exe, or downloads it via the official install.ps1 when none is found. No elevation required.
Local models panel ("Available") — lists models the running server reports via GET /models, enriched with catalog metadata (display name, params, size, brand logo, vision flag). Load/unload a model and open its WebUI from a row's action glyph.
Recommended Models panel — a scrollable list of downloadable builds (quants) from the Hub, with a one-click download that streams real-time progress over the server's GET /models/sse endpoint.
Settings window — Hugging Face access token (for gated/private repos) and a configurable GGUF cache directory, persisted per-user.
Per-user logging — daily-rotated logs at %LOCALAPPDATA%\LlamaApp\logs\LlamaApp-YYYYMMDD.log (7-day retention).

Solution layout

Project	Target	Role
`LlamaApp`	`net10.0-windows10.0.19041.0`	WinUI 3 unpackaged app: tray icon, flyout shell, settings window, views.
`LlamaApp.LlamaCpp`	`net10.0`	Detects/installs/launches the `llama` binary, talks to the server REST API, parses SSE download streams.
`LlamaApp.HuggingFace`	`net10.0`	Hugging Face Hub client and remote catalog fetcher (`https://llama.app/v1/catalog.json`).
`LlamaApp.Catalog`	`net10.0`	Generic catalog layer over `IModelSource` (abstraction shared by Hub + local sources).
`LlamaApp.Common`	`net10.0`	Shared interfaces (`IModel`, `IModelSource`), logging, `ModelDownloadProgress`.
`LlamaApp.LlamaCpp.Tests`	`net10.0`	xUnit tests for the server's JSON schemas and SSE parsing.

The helper libraries are pure managed (Any CPU); only the WinUI app builds for a concrete platform (x64/x86/ARM64).

Build & run

Requires the .NET 10 SDK and the Windows App SDK workloads (WinUI 3, WebView2).

dotnet restore
dotnet build -c Debug
dotnet run -c Debug

Build for a specific platform (required for the app to launch correctly):

dotnet build -c Release -r win-x64      # or win-x86 / win-arm64

Run the schema/SSE unit tests:

dotnet test LlamaApp.LlamaCpp.Tests

Note: Any-CPU builds are coerced to x64 in the app project so the x64 apphost matches the installed WebView2 runtime (the x64/ARM64EC WebView2 won't load under an ARM64-native apphost).

Runtime locations

What	Path
App-managed binary	`%USERPROFILE%\.llama-app\llama.exe`
GGUF model cache	`%USERPROFILE%\.cache\huggingface\hub` (shared with `llama.cpp` & HF tools)
Settings	`%LOCALAPPDATA%\LlamaApp\settings.json`
Logs	`%LOCALAPPDATA%\LlamaApp\logs\LlamaApp-YYYYMMDD.log`
Server	`http://localhost:2276` (`llama serve --port 2276`)

Architecture notes

LlamaManager.Shared is a singleton that owns the binary resolution + server process lifecycle, exposes StateChanged / ModelsChanged events, and is consumed by both App (startup) and MainWindow (live UI reconciliation).
The model-state poller fires ModelsChanged roughly every second with a fresh /models snapshot; the Available section reconciles rows in place (play → loading ring → open glyph) rather than rebuilding the list, to avoid flicker and lost click state.
SSE download progress is parsed from the server's GET /models/sse stream (ParseSseStreamAsync), which dispatches one JSON event per blank line (and flushes a trailing event at EOF).
The flyout dismisses itself on deactivation (focus loss) with short grace periods to avoid bounce-reopen quirks.

Llama-Windows

Specs

README