Smart Meet

Smart Meet – AI Meeting Copilot

A Chrome/Edge extension that listens to your online meetings with you. It flags questions that need an answer and points worth clarifying in a side panel while people talk. When the meeting ends, you can ask anything about the conversation. It runs on your own LLM API key, and there is no backend.

• Live insights. Questions, clarification points (vague deadlines, undefined acronyms, conflicting statements), action items, decisions and key facts. Items aimed at you are highlighted, and each comes with a suggested reply you can say out loud. A one-click “Draft answer” streams a fuller reply.
• Works anywhere.
  • Audio mode captures the meeting tab plus your mic on any site (Google Meet, Zoom web, Teams, Webex, Whereby…), with your voice labelled “You”.
  • Caption mode reads Meet, Zoom web or Teams web captions for free, with speaker names.
• Ask AI. Chat with any meeting, live or past. Answers cite [mm:ss] timestamps. Long meetings fall back to BM25 retrieval over the transcript.
• Summaries. Title, summary, decisions, action items (owner and due date) and open questions are written automatically when the meeting ends.
• History and export. Search past meetings and export them to Markdown or JSON. Automatic deletion is optional.
• Bring your own model. OpenAI, Anthropic Claude, Google Gemini, OpenRouter, Groq, Ollama (local), or any OpenAI-compatible endpoint. Speech-to-text runs on OpenAI, Groq, Deepgram, or a self-hosted Whisper server.
• Private by design. Data stays in IndexedDB on the device. Requests go straight from the browser to the provider you chose. There is no analytics or tracking. See PRIVACY.md.

Install

From a release: download smart-meet-<version>.zip from Releases and unzip it. Then follow steps 2–4 below, selecting the unzipped folder.

From source:

1. npm run icons (only needed if you change the icon; the PNGs are committed).
2. Open chrome://extensions (or edge://extensions) and turn on Developer mode.
3. Click Load unpacked and select this folder.
4. The settings page opens. Add an API key, pick a capture mode, then open a meeting, click the toolbar icon, and press Start.

Requires Chrome or Edge 116+. Firefox and Safari don’t support the side panel and tab-capture APIs this extension uses.

Quick start configurations

Goal	AI model	Transcription	Capture mode
One key, best quality	OpenAI gpt-5-mini	OpenAI gpt-4o-mini-transcribe	Automatic
Claude	Anthropic claude-opus-5-5 (optionally a faster live model such as claude-haiku-4-5)	Groq or OpenAI	Automatic
Cheapest	Groq llama-3.3-70b-versatile	Groq whisper-large-v3-turbo	Automatic
No speech-to-text bill	Any	None	Live captions (turn on CC in the meeting)
Fully local	Ollama	Custom → local Whisper server (e.g. speaches / faster-whisper-server)	Audio

For Ollama, start it with OLLAMA_ORIGINS=chrome-extension://* so the extension is allowed to call it.

How it works

 Meeting tab ──tabCapture──▶ offscreen document ──▶ speech-to-text API
      │                     (VAD-chunked webm/opus,     │
      │                      tab audio + mic)           ▼
      └─ caption content script ───────────────▶ service worker ──▶ IndexedDB
                                                  │  (session, transcript,
                                                  │   insight scheduling,
                                                  ▼   summary)
                                              LLM provider
                                                  ▲
 Side panel (Live · Ask AI · History) ────────────┘ streaming Q&A, draft replies

• src/background/service-worker.js: session lifecycle, transcript intake, debounced live analysis (it runs every N seconds or after M new characters, whichever comes first), and the summary at the end of the meeting.
• src/offscreen/offscreen.js and src/lib/recorder.js: record tab and mic audio, cut it into chunks at natural pauses, skip silent chunks, transcribe chunks in order, and filter out common Whisper hallucinations.
• src/content/captions.js: reads live captions in Meet, Teams and Zoom web, sending text once it has stopped changing and de-duplicating text that grows or gets revised (caption-diff.js).
• src/lib/providers.js: one complete / stream / listModels interface over every LLM provider, with retries, JSON mode and a fallback when an endpoint rejects JSON mode.
• src/lib/analyzer.js and src/lib/qa.js: prompts, schemas, parsing, de-duplication, retrieval and summarisation.
• src/sidepanel/, src/options/: the UI, written in plain ES modules with no framework and no build step.

Claude-specific behaviour

For Anthropic models, the extension calls the Messages API directly from the browser with the anthropic-dangerous-direct-browser-access header. It caches the transcript-bearing system prompt (cache_control) so follow-up questions are cheaper. Effort is set only on models that support it: low for live detection and draft replies, medium for Q&A and summaries. JSON comes from structured outputs (output_config.format).

On Claude Opus 5 / 5.5, Sonnet 5.5 and Fable 5.1, requests opt in to server-side refusal fallback (fallbacks: "default"), and a refusal stop reason produces a clear error message. To turn the fallback off, remove anthropicSupportsFallback in providers.js.

Development

npm test          # unit tests (node:test, no dependencies)
npm run check     # validates manifest + every file reference
npm run build     # check, then package dist/smart-meet-<version>.zip

There are no runtime or dev dependencies. CI (.github/workflows/ci.yml) runs the tests and uploads the store zip as a build artifact on every push. Pushing a v<version> tag runs .github/workflows/release.yml, which publishes the zip as a GitHub release.

Manual test checklist before a release

• Google Meet, audio mode: Start from the toolbar icon, confirm others are transcribed as “Others” and you as “You”, and that you can still hear the meeting.
• Google Meet, caption mode: turn on CC; speaker names appear; the “turn on captions” hint shows when CC is off.
• Zoom web client (app.zoom.us/wc) and Teams web in both modes.
• An insight appears within about 20 s of someone asking you a question; Draft answer streams; Done, dismiss and the badge count update.
• Stop: the summary appears in History; Ask AI answers with timestamps; Markdown and JSON export open correctly.
• Closing the meeting tab stops the capture; an invalid API key shows a banner rather than failing silently.
• Settings: Test connection works for each provider you support; Delete all clears History.

Deployment

Smart Meet has no backend. Deploying it means publishing the privacy policy on GitHub Pages, tagging a release, and submitting the zip to the Chrome Web Store and Microsoft Edge Add-ons. The full step-by-step guide is in docs/DEPLOYMENT.md.

In short:

1. Bump version in both manifest.json and package.json (npm run check enforces that they match), then update CHANGELOG.md.
2. Run npm run verify and go through the manual test checklist above.
3. Commit, then git tag v<version> and git push origin main v<version>. The Release workflow builds smart-meet-<version>.zip and attaches it to a GitHub release.
4. Upload the zip to the Chrome Web Store Developer Dashboard and Edge Partner Center. The listing text is in docs/STORE_LISTING.md, and the screenshots, promo tile and logo are in docs/store/.
5. Privacy policy URL for both stores: https://ammarkarimi.github.io/smart-meet/PRIVACY.html. Enable it once under Settings → Pages (main branch, root folder).

Limitations

• The Zoom and Teams desktop apps can’t be captured by a browser extension. Join from the browser instead.
• Caption selectors depend on each web client’s markup. If a vendor redesign breaks caption reading, audio mode still works. Selectors are listed at the top of src/content/captions.js.
• In audio mode, all remote participants are labelled “Others”. Use caption mode on Meet or Teams if you need per-speaker names.
• Chrome allows tab capture only after you invoke the extension on that tab. If Start reports this, click the toolbar icon on the meeting tab, then press Start.