Meeting transcripts

Zoom AI Companion transcripts land as markdown in a private S3 bucket, mirrored to a local folder agents grep.

Meeting transcripts

Every meeting becomes one markdown file an agent can read. No video, no database, no vendor.

The two rulings that shape everything else:

Fetch the AI Companion retained transcript, not the cloud recording one. No recording means no storage quota and no dependence on someone pressing Record.
A private S3 bucket is the system of record. A local mirror, outside any git repo, is a cache. That follows an existing standard — see Storage.

  ZOOM                     THE MAC                      DURABLE
  ┌──────────────┐
  │ AI Companion │
  │   summary    │  auto-start required
  │      ↓       │
  │  transcript  │
  │  retained    │
  └──────┬───────┘
         │  GET /meetings/meeting_summaries?from=&to=   ENUMERATE
         │  GET /meetings/<meeting-uuid>/transcript     FETCH (VTT)
         ▼
   ┌───────────────┐
   │ launchd, 06:00│  polls a 7-day overlapping window.
   │ + on wake     │  Missed a day? Next run heals it.
   └──────┬────────┘
          │  VTT ─▶ markdown + YAML frontmatter
          ▼
   ┌───────────────┐        aws s3 cp        ┌──────────────────┐
   │ ~/transcripts │ ─────────────────────▶  │ private S3 bucket│
   │  local mirror │ ◀───── aws s3 sync ──── │ SYSTEM OF RECORD │
   │  (cache)      │                         │ versioned, SSE   │
   └──────┬────────┘                         └──────────────────┘
          │
          ▼  Glob / Grep / Read — no creds, no network, no SDK
     ┌─────────┐
     │  agent  │
     └─────────┘

Where transcripts must never go: `docs/`

Not in this repo. Two reasons, both hard:

npm run validate caps a page at 500 words. A transcript is thousands.
This wiki renders publicly right now. Google sign-in is built but parked in unmerged PR #17. A client transcript here is a disclosure, not a typo.

The wiki holds the spec and the pointers. Never the text.

The pages

Page	What it settles
What Zoom gives you	Three artifacts, one of them right
Zoom settings	The toggles an admin must flip first
Zoom API	Endpoints, scopes, app type, gotchas
Pipeline	Poll, not webhooks. Enumerate then fetch
File shape	Naming, frontmatter, VTT cleanup
Storage	S3 is the record. The mirror is a cache
Growth path	What to build later, and the trigger for each
Consent	Internal vs client. Needs a human ruling
Open questions	Four decisions still owed

Scale: about 13.5 MB and 750 files a year. S3 bills under $0.01/month for that. Cost is not an argument for or against anything here.

Every example on these pages uses fake names — Sample Person, ACME Corp, <meeting-uuid>. Keep it that way. This wiki is public.

Meeting transcripts

Meeting transcripts

Where transcripts must never go: docs/

The pages

Where transcripts must never go: `docs/`