For the thread you keep putting off

Less scrolling.
Not more made-up words.

Paste a handoff, meeting notes, or a long thread. Open-weight AI picks a few useful sentences. You get the original quotes and their sources, not a generated story.

What can you skip reading twice?

40-16,000 characters; 3-80 distinct sentences of at least 12 characters. Prefer focused English-language notes, not credentials or complete chat archives.

UTF-8 text only, up to 48 KB. Opening a file replaces the input in this tab only.

The notes stay in this tab. First use downloads a public AI library and model from jsDelivr and Hugging Face. Inference runs on your device. No account, API key, analytics, or notes storage.

Ready. The model loads only when you request a digest.

What runs locally, and what does not?

Transformers.js 3.8.1 runs the quantized all-MiniLM-L6-v2 model on CPU through WebAssembly. A cold start downloads public assets, often tens of MB, so internet access is initially required. Once loaded, the open tab can run inference offline; a fresh reload is not guaranteed to work offline.

The model revision is pinned. Notes never go to an inference API or local storage. Public asset providers still see ordinary network metadata. An explicit digest download creates a local file containing your selected quotes.

Long sentences exceeding the model's 256-token limit are rejected with an actionable error rather than silently truncated. Identical sentences, ignoring whitespace and case, are deduplicated before inference. Sentences shorter than 12 characters are omitted and counted in the report.