Does this really run on my device?

Yes. The model is downloaded to your browser and runs on your own device's GPU. Your transcript never leaves your computer — nothing is uploaded to a server.

How is this different from the summarizer?

The summarizer writes a general summary in one pass. Meeting Notes runs a two-pass map-reduce pipeline tuned for meetings: it extracts structured facts from each section first, then synthesizes decisions, action items, topics and next steps — deduplicated and action-oriented.

Can I turn a recording into notes?

Yes — transcribe the audio or video first with the in-browser transcriber, then paste the transcript here. Both tools run entirely on your device.

Any meeting or call transcript as plain text, or a subtitle file (.srt / .vtt). Long transcripts are handled automatically by analyzing them in sections and combining the results.

Why does the first run take longer?

The first run downloads the AI model (about 0.9 GB), then it's cached in your browser, so later runs start quickly. A desktop with WebGPU is required for this tool.

Which model does it use?

Qwen2.5-1.5B-Instruct (Apache-2.0) via WebLLM/MLC, executed on WebGPU. A browser without WebGPU cannot run it, so it says so up front rather than falling back to something that would take minutes.

Meeting notes from a transcript

Paste a meeting transcript and get a clean, structured report: decisions, action items with owners, discussion topics, unresolved questions and next steps. It runs entirely on your device — nothing is uploaded, so it's free, unlimited, and private. Have a recording? Transcribe it first, then bring the transcript here.

How to turn a transcript into meeting notes

Paste your transcript

Paste a meeting or call transcript, or drop a .txt, .srt or .vtt file. Speaker names and blank lines help the AI find structure. It never leaves your browser.
Pick a report style

Choose a full structured report, a short TL;DR, or category-tagged key points — whatever fits how you'll use the notes.
Generate & save

The AI runs on your device in two passes — it analyzes each section, then writes the report (first run downloads the model, then it's cached). Copy it or download the .md.

Why generate them here

Transcript to action items

A two-pass analyze-then-synthesize pipeline pulls out decisions and action items with owners and due dates — not a vague summary, but a report you can act on.

Your files are never uploaded

It all runs 100% in your browser — no server, no account, no upload. Ideal for confidential calls, standups and client meetings.

Free & unlimited

Because inference happens on your device, there's no per-token cost and no cap. Generate notes for as many meetings as you like.

The model behind it

Powered by Qwen2.5-1.5B-Instruct, an Apache-2.0 language model, run in your browser through WebLLM on WebGPU. Meeting transcripts are among the most sensitive text people handle, and this one never leaves the machine it was typed on.

Questions

Does this really run on my device?: Yes. The model is downloaded to your browser and runs on your own device's GPU. Your transcript never leaves your computer — nothing is uploaded to a server.
How is this different from the summarizer?: The summarizer writes a general summary in one pass. Meeting Notes runs a two-pass map-reduce pipeline tuned for meetings: it extracts structured facts from each section first, then synthesizes decisions, action items, topics and next steps — deduplicated and action-oriented.
Can I turn a recording into notes?: Yes — transcribe the audio or video first with the in-browser transcriber, then paste the transcript here. Both tools run entirely on your device.
What can I paste?: Any meeting or call transcript as plain text, or a subtitle file (.srt / .vtt). Long transcripts are handled automatically by analyzing them in sections and combining the results.
Why does the first run take longer?: The first run downloads the AI model (about 0.9 GB), then it's cached in your browser, so later runs start quickly. A desktop with WebGPU is required for this tool.
Which model does it use?: Qwen2.5-1.5B-Instruct (Apache-2.0) via WebLLM/MLC, executed on WebGPU. A browser without WebGPU cannot run it, so it says so up front rather than falling back to something that would take minutes.