meet2linear

Turns Google Meet calls into Linear tickets — entirely inside a Chrome extension, no server, no paid services:

1. A content script scrapes Meet's free live captions (speaker-attributed) during the call.
2. When the call ends, the transcript lands in chrome.storage and Gemini Flash (free-tier API key) extracts candidate tickets — bugs, feature requests — with verbatim evidence quotes.
3. Click the toolbar icon to open the review page: edit, approve or discard candidates; approved ones are created in Linear (free-plan GraphQL API) with a from-meet label.

Node is needed only to build the extension; nothing runs outside the browser. Gemini and Linear are called directly from extension contexts (raw REST/GraphQL — no SDKs bundled).

Caption-capture mechanics (selectors, ASR-revision handling, end detection) are adapted from transcriptonic.

Layout

• shared/ — types + Zod schemas
• extension/src/
  • content.ts — caption capture on meet.google.com; selectors.ts holds every Meet DOM selector (the churn blast-radius)
  • background.ts — service worker: single writer of meeting records, auto-runs extraction at meeting end
  • store.ts — chrome.storage-backed meetings + settings
  • extraction.ts / prompt.ts — Gemini structured-output extraction (gemini-flash-latest)
  • linear.ts — raw GraphQL: teams, label bootstrap, issue creation
  • review/ — the review page (toolbar icon opens it)

Setup

npm install
npm run build          # or: npm run watch

Chrome → chrome://extensions → enable Developer mode → Load unpacked → select extension/dist/.

Then click the extension's toolbar icon → ⚙ Settings:

1. Paste a Gemini API key (free from Google AI Studio).
2. Paste a Linear personal API key (Linear → Settings → Security & Access; free plan works).
3. Click Load teams and pick the team issues should go to.

Keys live in chrome.storage.local on your machine.

Try it without a meeting

On the review page, click Load sample call — it imports a bundled fake customer call and runs extraction. The fixture contains three bugs (one mentioned twice — tests dedup), one feature request, and two red herrings (an issue resolved live on the call, and small talk). Approve one to verify Linear creation end-to-end.

End-to-end checklist

1. Extension built, loaded unpacked, keys + team set in Settings.
2. Start a Meet with yourself (meet.google.com → New meeting); wait for the "capturing captions" banner (the extension auto-enables CC; if it can't find the button it tells you to enable CC manually).
3. Say a few scripted sentences aloud, e.g. "The CSV export is broken, the header row is missing" — captions should appear as you speak.
4. Leave the call — extraction runs automatically.
5. Click the toolbar icon: the meeting is listed, candidates appear within seconds.
6. Edit a title if you like, Approve → Linear, follow the issue link.

Notes & limitations

• Captions must be on — the extension auto-clicks CC and warns on-page if it can't.
• Meet DOM churn is the main fragility. All selectors are in extension/src/selectors.ts, favoring ARIA attributes and icon-ligature names over class names. If capture silently stops, start there (a health monitor shows an on-page banner when the caption region disappears).
• Evidence quotes the model didn't copy verbatim are flagged ⚠ in the review UI (hallucination guard).
• Re-extraction replaces only proposed candidates; approved/discarded ones are kept.
• The service worker is the sole writer of meeting records; the review page updates live via storage.onChanged.
• Free-tier limits are a non-issue at this volume: one Gemini call per meeting; Linear calls only on approve.
• History: git log has an earlier iteration where extraction/Linear ran on a local Node server instead of in the extension.