What is the Meeting Note-Taker & Summarizer?

It is a privacy-first tool that records or imports meeting audio, transcribes it into a timestamped transcript, and generates structured meeting minutes - a summary plus decisions, action items, and open questions. Everything runs entirely in your browser with no backend.

Is my meeting data private?

Yes, completely. All audio capture, transcription, and summarization happen locally in your browser using Web Workers. No recordings, transcripts, or minutes are ever uploaded to any server, which makes it safe for confidential meetings.

What audio sources are supported?

You can record from your microphone (captured as the [Self] channel), optionally add tab or system audio (for example a video call) via screen sharing as a separate [Remote] channel, upload an existing audio file (.wav, .mp3, .m4a, .webm), or paste and upload a text transcript (.txt or .md). This means you can use the tool even without a live microphone.

Which AI models are used?

Transcription uses OpenAI Whisper (whisper-base) via Transformers.js, with Moonshine (moonshine-base) available as a fast short-turn model. Minute generation uses Qwen2.5-0.5B-Instruct through an engine chosen for your hardware, and all three engines are grammar-constrained to emit schema-valid JSON: WebGPU-capable devices with ample memory run WebLLM (@mlc-ai/web-llm) with XGrammar; constrained WebGPU devices run the Transformers.js map-reduce summarizer, grammar-constrained with @huggingface/transformers-structured-output; and devices without WebGPU run the wllama CPU/WASM engine (@wllama/wllama, Qwen2.5-0.5B GGUF via single-thread llama.cpp Wasm) with GBNF for the same schema-valid output on the CPU. If you opt in to speaker identification, a diarization pass additionally loads Silero VAD, pyannote segmentation, and WeSpeaker embedding models to group speakers. If you opt in to vision ingestion, Microsoft Florence-2 (OCR/layout) and optionally SmolVLM2 read on-screen text from a shared screen or tab. All models run with WebGPU acceleration and automatic WASM fallback. A lightweight, deterministic keyword tagger (not an AI model) flags likely decisions, action items, and open questions in the transcript.

Can it read slides or screen-share text (vision)?

Yes, as an opt-in extra. Toggle 'Capture slides / screen share (vision)' and, when you start recording, pick a screen or tab to share. The tool keeps the video track and samples frames, uses a perceptual hash to keep only visually-distinct keyframes (so repeated/unchanged frames never trigger a model), and reads the on-screen text locally with Microsoft Florence-2 (OCR and layout). You can optionally add a SmolVLM2 caption of each keyframe. The extracted text is shown as keyframe thumbnails and bound into the on-device knowledge graph as VisualArtifact nodes. This is opt-in and heavy: the vision models are multi-GB and only download when you enable the feature, so use WiFi. Screen-frame capture relies on MediaStreamTrackProcessor and OffscreenCanvas, which today only Chromium-based browsers (Chrome, Edge) provide; in other browsers the toggle is disabled with a note and audio transcription still works. Everything stays on your device.

How accurate are the transcript and minutes?

The models are small so they can run on-device, which means transcription and summaries are good but not perfect. Expect occasional misheard words and paraphrased minutes. Always review the generated decisions and action items against the transcript before sharing.

Why is the first load slow?

On first use the core AI models need to download (roughly 200-400MB total depending on your device). After that, models are cached in your browser so subsequent visits load quickly. Optional features download more only when enabled - speaker diarization adds a few more models, and vision ingestion adds the multi-GB Florence-2 (and optional SmolVLM2) models. Use WiFi for the first download to avoid mobile data charges.

Can it tell speakers apart?

Partly. Recording keeps your mic ([Self]) and any tab/system audio ([Remote]) as two separate labeled channels instead of mixing them. You can also opt in to speaker diarization, which groups the [Remote] channel into Speaker 1, Speaker 2, and so on using voice-activity detection, segmentation, speaker embeddings, and clustering. This is best-effort: overlapping speech, short turns, similar voices, or noisy audio can merge or split speakers, and it does not identify people by name. Imported audio files are labeled 'Participant'.

Can I search back through the transcript?

Yes. Once a transcript exists, the Search Transcript box runs a local hybrid search over 15-30s speaker turns, combining BM25 keyword matching with vector (semantic) similarity from bge-small-en-v1.5 embeddings in an Orama index. It returns the most relevant, timestamped, speaker-attributed passages, and clicking a result scrolls straight to that moment in the transcript. The search runs entirely in your browser; to keep it fast and lightweight each turn is also stored as an int8 and a 1-bit (sign) quantized vector used as a cheap prefilter.

Does it support agent tool-calling (WebMCP)?

Yes, as an additive agent mode. The on-device System-2 model can emit grammar-constrained calls to five meeting tools (record a decision, upsert an action item, query the transcript, update a topic summary, synthesize topic minutes). These calls run through an actual Model Context Protocol (MCP) server that lives entirely in your browser - built from the @modelcontextprotocol/sdk protocol core over a custom MessagePort transport, with no network and no stdio. Tool execution is handled by pure, unit-tested reducers, so the agent and the regular pipeline produce the same kind of minutes. This is progressive enhancement: the standard transcribe-tag-summarize-export flow works exactly the same when agent mode is not driving.

Are my decisions and action items saved between sessions?

In agent mode, yes - on your device. Decisions and action items the agent records are persisted into a local SQLite knowledge graph stored in your browser's Origin Private File System (OPFS), using the SQLite-WASM opfs-sahpool storage engine, so they survive a page reload in that tab. Nothing is uploaded. Because this storage uses an exclusive lock, only one tab of the site can hold the durable graph at a time; if durable storage is unavailable (an unsupported browser, no cross-origin isolation, or the lock is held by another tab) the tool automatically falls back to in-memory storage for the session - the badge in the Knowledge graph panel reads 'OPFS (durable)' or 'Memory only' so you always know which applies. Memory-only data is cleared when you reload. Updates to an existing decision or action item create a new version and supersede the old one (the full history is kept), and a crash-recovery write-ahead log replays any unfinished writes when the tool starts. Older records follow a 90-day, 3-tier (hot/warm/cold) retention and compaction policy.

What are the current limitations?

The tool works best in English and transcribes after you stop recording rather than fully live. Speaker diarization is opt-in and best-effort rather than exact. Vision ingestion is opt-in, pulls multi-GB models, and the screen-frame capture pipeline is Chromium-only (the toggle is disabled in other browsers). On-device persistence is single-tab (the durable OPFS store cannot be shared across tabs at the same time). Additional languages are a possible future addition.

AI Meeting Note-Taker

Record or import meetings and get an AI transcript plus structured minutes with decisions & action items, persisted on-device in a local knowledge graph - 100% private, runs in your browser

ai meeting note-taker

meeting transcription

meeting summarizer

Record Meeting

Include tab / system audio as a separate [Remote] channel (shares a tab or screen; video is discarded)

Identify speakers (diarization) on the [Remote] channel

Capture slides / screen share (vision): read on-screen text from a shared tab or screen

Your browser will ask for microphone permission. Your mic is captured as the [Self] channel and any tab/system audio as a separate [Remote] channel. Audio is transcribed locally after you stop recording.

Import Audio or Transcript

Upload an audio file (.wav, .mp3, .m4a, .webm) or a transcript (.txt, .md)

Or paste a transcript (plain lines, or "[MM:SS] Speaker: text")

Transcript

No transcript yet. Record a meeting, upload audio, or import a transcript to begin.

Meeting Minutes

Loading Meeting Note-Taker...

How to Use

Click Start Recording and grant microphone permission.
Optionally toggle "Include tab / system audio" to capture a shared tab or call as a separate [Remote] channel.
Optionally toggle "Identify speakers (diarization)" to group the [Remote] channel into Speaker 1..N (opt-in, best-effort, downloads extra models).
Optionally toggle "Capture slides / screen share (vision)" to read on-screen text with Florence-2 (opt-in, heavy multi-GB models, Chromium-only).
Stop recording to transcribe, or import an existing audio file or transcript instead.
Review the timestamped transcript (your mic is labeled [Self], tab/system audio [Remote], imports [Participant]).
Use the Search Transcript box to find any moment by meaning or keyword (hybrid BM25 + vector search), then click a result to jump to it.
Click Generate Minutes to produce a summary, decisions, action items, and open questions.
Export your notes as Markdown or JSON.

How it Works

The entire pipeline runs on-device in your browser using Web Workers, WebGPU (with a WebAssembly fallback), and AI models from Transformers.js:

1. Dual-Channel Capture

Your mic ([Self]) and optional tab/system audio ([Remote]) are captured as two separate channels through an AudioWorklet and resampled to 16kHz mono PCM.

2. Whisper / Moonshine ASR

Each channel is transcribed into timestamped utterances with Whisper (Moonshine is available for fast short turns).

3. Optional Diarization

If enabled, a best-effort pass (voice-activity detection, segmentation, speaker embeddings, and cosine clustering) groups the [Remote] channel into Speaker 1..N labels.

4. Optional Vision Ingestion

If enabled, a shared screen/tab's video is sampled, a perceptual hash keeps only visually-distinct keyframes, and Florence-2 (with an optional SmolVLM2 caption) reads the on-screen text locally. Keyframes are shown as thumbnails and bound into the knowledge graph as VisualArtifact nodes. Opt-in and heavy (multi-GB models), Chromium-only.

5. Candidate Tagging

A deterministic tagger flags utterances that look like decisions, action items, or open questions, shown inline in the transcript.

6. Hybrid Transcript Search

15-30s speaker turns are embedded with bge-small-en-v1.5 into an Orama index. A hybrid BM25 + vector search finds the right moment by meaning or keyword; int8 and 1-bit quantized vectors keep it fast and lightweight.

7. Minutes (hardware-tier engine)

Qwen2.5-0.5B turns the transcript into structured minutes, and all three hardware-tier engines are grammar-constrained to emit schema-valid JSON. On WebGPU-capable devices with ample memory it runs WebLLM with XGrammar grammar-constrained decoding; constrained WebGPU devices use the Transformers.js map-reduce summarizer, grammar-constrained with @huggingface/transformers-structured-output (a StructuredOutputProcessor applied to both the map and reduce passes); and devices without WebGPU use the wllama CPU/WASM engine (Qwen2.5-0.5B GGUF via single-thread llama.cpp Wasm) with GBNF grammar-constrained decoding. The active engine and tier are shown in the badges.

8. WebMCP Tool-Calling (agent mode)

Optionally, the on-device model can act as an agent: it emits grammar-constrained calls to five meeting tools (record a decision, upsert an action item, query the transcript, update or synthesize a topic summary). The calls run through an actual Model Context Protocol (MCP) server that lives entirely in your browser over a MessagePort - no network, no stdio. The deterministic pipeline still works when agent mode is off.

9. On-Device Knowledge Graph

In agent mode, decisions and action items are saved into a local SQLite knowledge graph in your browser's Origin Private File System (OPFS), so they survive a reload in this tab. When vision ingestion is on, captured keyframes are bound in as VisualArtifact nodes (with DEPICTS and MODIFIES_VISUAL_STATE edges). If durable storage is unavailable (an unsupported browser, or the storage lock is held by another tab), it falls back to memory-only and the badge says so. Updates supersede older versions instead of overwriting, and a crash-recovery write-ahead log replays unfinished writes on startup.

10. Structured Output

You get a summary plus decisions, action items, and open questions.

11. Export

Download everything as Markdown or JSON for sharing or archiving.

Privacy & Security

Everything happens entirely in your browser. No audio, transcript, or minutes are ever uploaded to a server.

• Audio is processed locally using Web Workers
• AI models run on-device via WebGPU or WASM
• No recordings or notes leave your browser at any point
• Models are cached locally after the first download
• Persisted decisions and action items are stored only on your device (local SQLite in OPFS), never uploaded, and stay in a single tab
• Optional vision ingestion is opt-in and local: captured screen frames are read on-device with Florence-2 and never uploaded
• Ideal for confidential meetings and sensitive discussions

Frequently Asked Questions