> Documentation index: https://gethologram.ai/llms.txt

# Apps

> Apps that run entirely in your browser, each in its own sealed frame, with every model byte verified before it is used.

An App is an app and the model it needs, sealed under one address and run by the visitor's own browser. Nothing runs on a server: the page is a folder of static files, the model comes from the host the App names, and the work happens on the visitor's GPU. The hub lists five at [/spaces/](https://gethologram.ai/spaces/): speech, image, chat, vision and depth.

## What is sealed

Every file in an App's folder is named in `holospace.lock.json` with its SHA-256; the runtime files it depends on (the verified fetch, the store, its ONNX Runtime build) are named there too. The lock's `root` is the SHA-256 over that map. Change one byte anywhere and the root changes. The catalog at [/spaces/spaces.json](https://gethologram.ai/spaces/spaces.json) carries each root.

## How the model is verified

The App names its model (`onnx-community/Kokoro-82M-v1.0-ONNX`, say). Before the app's code runs, a small prelude reads the model index — `/api/models/{owner}/{name}/tree/main` on the model host — and from then on every fetch of a model file is held until its bytes re-derive to the digest the index gave for that path: SHA-256 for LFS files, git's blob SHA-1 for the small ones. A byte that does not match is refused; the app sees a failed load, never a wrong file. A verified file is kept in the App's own browser store by its digest, so the second open needs no network.

The host is a location, not the identity. Today the demo Apps read from huggingface.co; when the hub indexes the same models, the same Apps read from here with the same digests.

## Isolation

Each App runs in a sandboxed frame on this origin with its own storage namespace, and a Content-Security-Policy written into its page allows exactly three things: its own files, the hub's brand kit, and the one model host it declares. No CDN script, no third-party asset, no analytics. The frame asks for no camera or microphone.

## On the registry

A published App is an OCI artifact at `/v2/spaces/<id>`: config `holospace.json`, one layer per file, artifact type `application/vnd.hologram.space.v1+json`, the sealed root in an annotation, tag `latest`. Reads need no token. The page asks the registry for each App's manifest and marks the ones it finds.

```bash
curl -sI https://gethologram.ai/v2/spaces/kokoro-tts/manifests/latest \
  -H "Accept: application/vnd.oci.image.manifest.v1+json" | grep -i docker-content-digest
```

## What it needs

WebGPU and private file storage (OPFS): current Chrome, Edge or Safari on a device with a GPU. The page says so on the card when a browser lacks either. Model downloads are 176–390 MB on first open and come from the store after that; the first result waits on the GPU compiling the model, which can take a minute on integrated graphics.