Browser vs Desktop¶

The same codebase runs in both. What differs is how it reaches the outside world: files, processes, and GPUs.

	Browser	Desktop
Install	None — app.sleap.ai	~100–220 MB depending on platform, installer
Open / save `.slp`	File System Access API	Native file dialogs, writes in place
Drag-and-drop	✅	✅
Labeling, tracks, view, analyze	✅	✅
Video playback (WebCodecs)	✅	✅
Transcode legacy codecs	❌ convert externally	✅ native ffmpeg sidecar, cached
Local training / inference	❌	✅ via Environment
Remote training / inference	✅	✅
Crash-recovery drafts	✅ OPFS	✅ app data directory
Reveal file in file manager	❌	✅
Open Preferences Directory / Clear Video Transcode Cache	❌	✅
Local scrub proxies for network videos	❌	✅
NWB export	❌	✅
ONNX / TensorRT model export	❌	✅
Model-output overlays	❌	✅
DeepLabCut import	Chromium browsers only (Chrome, Edge) — needs the folder picker	✅
In-app updates	Reload the page	✅ per channel
Offline	❌ needs a connection to load the page	✅

Which should I use?¶

Browser if you are labeling, reviewing predictions, or checking someone else's project — there is nothing to install and nothing to keep up to date. It is also the right choice on a machine you don't administer, and it can still train and predict through a remote worker.

Desktop if you want native file handling on large projects, a local GPU, or legacy-codec videos handled for you.

Projects are the same .slp files either way, so moving between them costs nothing.

Saving differences worth knowing¶

In the browser, opening a file through the picker gives the page a writable handle and Save writes back to that same file. Opening by drag-and-drop yields no handle, so Save falls back to a Save-As prompt. On the desktop, Save always writes to the project's path.

See Saving & Recovery.

Under the hood¶

The desktop app is Tauri v2 — it uses the operating system's own WebView rather than bundling a browser engine. It does bundle its own ffmpeg/ffprobe for converting legacy-codec videos. Platform-specific behavior is isolated behind a small abstraction layer, so features arrive in both runtimes at once unless they fundamentally cannot.