GenClass

The version of Jev that runs in your browser. Fully open source (Apache-2.0).

GenClass is a small, fast typed-decision model and a Chrome extension built on it. You speak, and your browser acts, often before you finish the sentence. Everything runs locally: the model executes in the browser with WebGPU/WASM, and speech recognition uses Moonshine on-device by default.

GenClass is an independent open-source project. It is not affiliated with or endorsed by TypeSafe AI. "Jev" is TypeSafe's product name, used here only to describe what GenClass is comparable to.

What it does

• Insanely fast browser use. "Open GitHub and search for whisper": the open fires mid-sentence and the typing follows as soon as you finish the phrase.
• Ads and content filtering. Classify page blocks as ad, sponsored, clickbait or off-topic, then hide them.
• Focus modes. Tell it what you're working on, and it closes or parks tabs that don't fit.
• RAM management. Chrome eating all your memory? GenClass discards idle and irrelevant tabs before your machine starts swapping.
• Video and more (planned). Any decision that is a choice, yes/no or score over text runs through the same model.

GenClass vs Jev (direct, same inputs)

Decision	Jev	GenClass
Recognise the command mid-sentence	66.4%	90.4%
Pick the exact words to type	75.5%	94.9%
Full-command action	91.4%	92.0%
Pick the on-screen element	86.6%	82.4%
General held-out questions	94.7%	80.5%
Option-order flips	10–13% (third-party)	0
Cost / privacy	paid cloud API	free, on-device

Methodology, per-question tables and caveats are in BENCHMARKS.md. In short, GenClass wins at computer control and Jev wins at general questions.

How it works

GenClass answers typed questions about a state in one forward pass:

• choice over options;
• noul for yes/no;
• score over ordered levels.

The questions use the same request format as Jev's System One API. The browser layer turns each partial transcript plus the visible page elements into one request. It acts on closed-set commands as soon as they are complete, and it asks for confirmation before anything risky.

Open source: everything is here

• The model. A ModernBERT/Ettin encoder with typed decision heads, and per-option attention isolation so the order of the options can't change the answer (jev_local/engine/encoder/).
• A Jev-compatible API server. POST /v1/systemone, the same request and response shapes, so existing Jev clients work by changing the base URL (jev_local/server/).
• The voice computer-use harness. It re-decides on every partial transcript and acts mid-sentence (jev_local/harness/, docs/DEMO.md).
• Training and data pipelines. Synthetic data, multi-node CPU training and decontamination (jev_local/train/, jev_local/data/).
• The benchmark harness. 549 published Jev numbers, adapters for each publisher's own test code, and a pre-registration file (jev_local/bench/, bench/).
• Design docs and results. docs/, results/.

git clone https://github.com/MeharPro/GenClass && cd GenClass
python3.12 -m venv .venv && .venv/bin/pip install -e ".[dev]"
.venv/bin/python -m jev_local.server.app --ckpt <weights dir>   # Jev-compatible API on 127.0.0.1:8765

The model weights (Apache-2.0) and the Chrome extension zip are in the v0.1.0 release. The extension source is in extension/.

Install the Chrome extension (v0.1.0)

1. Download genclass-0.1.0.zip from the release and unzip it.
2. Open chrome://extensions, turn on Developer mode, click Load unpacked and pick the unzipped folder.
3. Click the GenClass icon to open the side panel. The first run downloads the model (~57 MB) and Moonshine speech (~63 MB) once, and both are cached. It starts in dry-run mode; switch on Live in Settings.

A Chrome Web Store listing is under review.

Status

Under active development. The extension, model weights (Apache-2.0) and install instructions are coming in the first release.

License

Apache-2.0