Synced from Hive. This page is pulled from hivecommons/hive@v5 during the docs build. Edit the canonical source in the Hive repository.

Betty on Zoom without a Zoom app: dial-in, desktop companion, live stream

Status: design only. Recommendation accepted in #18596; nothing on this page is built yet. The operator page for meetings as they work today is meetings.md; read it first. Related: meetings #18284, Betty #18391.

1. The wall

Betty cannot join Zoom as a web-client bot any more. The Zoom web client answers a Vexa bot with “automated bots aren’t allowed … must use Zoom RTMS”, and Zoom’s developer documentation now says the Meeting SDK “is reserved for human use cases and does not support bots or AI notetakers”. From March 2, 2026 an app that joins meetings outside its own account must be authorized (OBF or ZAK tokens, or RTMS).

We accept that wall. This page compares the ways Betty can still take part in a Zoom meeting:

  • without a Zoom Marketplace app or RTMS, and
  • without evading Zoom’s bot detection: no stealth browsers, no fingerprint spoofing, no anti-bot bypass of any kind.

Every option has to keep the three rules the meetings design already lives by:

  1. Consent first. Nothing from Betty before the consent notice, and the notice must be delivered, or Betty leaves.
  2. Hive is the policy point. Voice rules, chat rate limits, the research cap and the LLM budget are hive’s checks, whatever the platform.
  3. Transcripts are erased upstream after hand-off to the Inbox.

2. Options

2.1 Phone dial-in (PSTN through a telephony provider)

Betty places an outbound call to the meeting’s dial-in number, sends <meeting id>#, # (no participant id) and <passcode># as DTMF, streams the call audio to the shared transcription service and plays Kokoro TTS back into the call. A Quorum phone platform would own the call.

  • Account tier. Telephone dial-in is a paid-plan feature: Zoom’s free-plan page lists toll dial-in among what a paid upgrade adds, and Zoom Community moderators confirm the host must be licensed. Betty could phone into meetings hosted on paid accounts, never into meetings the owner hosts on a free (Basic) account, which are also capped at 40 minutes.
  • SIP instead of PSTN needs Zoom’s Conference Room Connector (H.323/SIP), a paid per-account add-on; it does not help a free account either.
  • Cost (Twilio US pay-as-you-go, October 2026 price page): outbound call $0.0140/min plus Media Streams $0.0044/min, $0.0184/min or about $1.10 per meeting-hour, plus $1.15/month for a number (or a verified caller id). Telnyx and other SIP trunks are in the same range.
  • Compliance. A phone participant is an ordinary, visible participant (shown as a phone number; it waits in the waiting room unless the host lets dial-ins bypass it). Zoom’s Terms do not single out phone participants, but an automated phone participant is close to what Zoom’s bot rules aim at, so Betty would only dial in when the host invited her.
  • Consent. A phone line has no chat, so the notice must be spoken on entry. Hive can confirm the audio played, not that anyone heard it: weaker than the confirmed chat delivery rule 1 asks for.
  • UX. Audio only, one mixed stream (no speaker names), DTMF fragile if the dial-in prompts change. Betty can speak.
  • Effort. Medium: a Media Streams websocket bridge in Quorum, DTMF sequencing, spoken consent, one new secret.
  • Needs from the owner. A telephony account and number, billing, and meetings hosted on a paid plan. Not startable today.

A small desktop app on the computer of the host (or any invited human) captures their own Zoom client’s audio and streams it to their hive as a live meeting. The pattern is the one Plaud Desktop ships: the app notices a Zoom, Teams or Meet call starting, asks “Let Betty listen to this call?” with the consent reminder, captures the call app’s audio (everyone else) and the microphone (the user), and streams both. Betty’s existing pipeline (transcript, Spek notes, research, replies) consumes it unchanged. Replies show in a small companion window and in the hive’s Meetings panel and Inbox. Betty speaking into the call through a virtual microphone is out of scope for v1.

  • Capture without a virtual driver. On macOS, ScreenCaptureKit captures one application’s audio (SCStreamConfiguration.capturesAudio, excludesCurrentProcessAudio; the microphone through captureMicrophone on macOS 15 and later), and Core Audio process taps (CATapDescription, AudioHardwareCreateProcessTap, macOS 14.2 and later, with an NSAudioCaptureUsageDescription) tap a single process. Windows has WASAPI per-process loopback; Linux has PipeWire per-node capture. BlackHole or VB-Cable are needed only for the later “Betty speaks through the user’s microphone” step.
  • Compliance. Nothing joins the meeting and nothing touches Zoom’s servers or bot detection: it is the user’s own client audio on the user’s own machine, like Plaud Desktop or Zoom’s own local recording. The legal duty is the one Zoom’s Terms (section 7) put on every user: comply with recording-consent laws.
  • Consent. Zoom does not notify participants, so the user’s attestation is the consent gate:
    • every call needs an explicit “I have told the participants” attestation, versioned like the Inbox recorder’s consent_version (inbox-recording.md);
    • hive records who attested, when, and which notice version;
    • a persistent red “Betty is listening” indicator shows while capturing, and one click stops it;
    • the companion offers the consent notice on the clipboard to paste into the meeting chat.
  • Privacy and security.
    • Audio leaves the machine only while a capture the user accepted is running, only to the user’s own hive, over TLS, as raw 16 kHz mono PCM.
    • The companion authenticates with a per-user, revocable companion token minted from the dashboard by an owner or read-write user. Only its SHA-256 is stored, as for the phone-widget token (src/pkg/dashboard/widget_token.go). Unlike the widget token, which is one per hive and read-only, the companion token belongs to one user and acts as that user at their live role: a revoked or downgraded user stops working at once, and read-only users cannot stream.
    • The token is admitted on the companion routes only and is carried in its own header, never in Authorization, so a hosted hive’s hub never takes it for a GitHub token.
    • Nothing is stored on the laptop beyond the token, kept in the OS keychain.
  • UX. Works on the free Zoom account and on every platform the user’s client joins (Zoom, Teams, Meet, Webex, Slack huddles). Two channels give “you” versus “the call” attribution; the call side is one mixed stream, with no per-person names. Without headphones the microphone picks up the speakers: v1 documents headphones and drops microphone segments that duplicate a call segment.
  • Cost. None: it uses the existing transcription service.
  • Effort. Medium-high: a desktop app, then one capture module per operating system.

2.3 Zoom “Live on Custom Live Streaming Service” (RTMP)

The host clicks More, Live on Custom Live Streaming Service, and Zoom pushes the meeting to an RTMP URL that Betty transcribes.

  • Account tier. Needs a Pro, Business, Education or Enterprise host and the admin settings “Allow livestreaming of meetings” and “Custom Live Streaming Service”. Not available on the free account.
  • Compliance. Fully sanctioned and visible: Zoom tells everyone the meeting is live.
  • Consent and replies. Listen-only: no chat and no voice, so Betty can neither post the consent notice nor answer in the meeting.
  • UX and effort. The host starts it in every meeting; 10 to 30 seconds of stream latency; the hive needs a public RTMP ingest endpoint, a new exposed surface. Cost: a Zoom Pro licence plus ingest bandwidth.

2.4 Other documented paths

  • Zoom RTMS or a Marketplace app. Zoom’s sanctioned route for AI notetakers; excluded by the brief because it needs a Zoom app.
  • Zoom cloud recording and transcript, then import. A paid feature, available only after the meeting; the existing Zoom cloud-recordings connector (meeting-recordings.md) already brings those into the Inbox.

3. Comparison

Phone dial-inDesktop companionRTMP live stream
Works with the owner’s free Zoom accountNo (host must be paid)YesNo (Pro or higher)
Close to Zoom’s bot rulesYes (automated participant)No (user’s own client)No (Zoom feature)
Consent deliverySpoken only, unconfirmablePer-call attestation, indicator, clipboard noticeNone in the meeting
Live Betty answersYes, by voiceYes, in the companion window and panel (voice later)Transcript only, 10 to 30 s late
Speaker attributionNoneYou versus the callNone
CostAbout $1.10 per meeting-hour plus $1.15/month$0Zoom Pro plus ingest
Needs an owner purchaseYes (telephony account)NoYes (Zoom Pro)
Other platforms at no costAny with dial-inZoom, Teams, Meet, Webex, SlackZoom only

4. Recommendation

Build the desktop companion (2.2), macOS first. It is the only option that works with a free Zoom account, needs no purchase, stays clear of Zoom’s bot rules because nothing joins the meeting, and reuses Betty’s live pipeline as is.

Keep phone dial-in (2.1) as a later add-on for meetings hosted on paid accounts, if the owner wants Betty to speak without the user’s machine; it needs a telephony account and spoken-consent rules first. Drop the RTMP live stream (2.3): listen-only, paid, and a new public ingest.

5. How it fits what hive already has

The dashboard’s ● Record (inbox-recording.md) already captures the microphone plus a meeting tab’s audio in Chromium through getDisplayMedia, behind a versioned consent notice, for owners and read-write users only. Its ingest (POST /api/inbox/recordings, ordered chunks?seq=N, finalize) is post-meeting: chunks must be WebM or MP4 media, they are concatenated at finalize and transcribed afterwards, so Betty cannot react during the call.

The companion keeps Record’s consent model and the shape of its chunk protocol (ordered seq, 409 with expected_seq on a gap, an idempotent resend) but streams to a live meeting route, so Betty answers during the call. When the meeting ends it lands in the Inbox as an ordinary meeting item, like a Vexa or Discord meeting, so no separate Record upload is needed.

6. Streaming contract (proposed)

The contract is to be fixed in the Quorum contract (docs/contract.md in hivecommons/quorum, a new companion platform) and in meetings.md when it ships. Companion to hive:

RouteBodyAnswer
POST /api/companion/meetings{title, app, consent_version, consent_attested: true}Creates a live meeting on platform companion (hive calls Quorum meeting.join with platform: "companion"), records who attested and when; returns {id, sample_rate: 16000, max_chunk_bytes}. A stale consent_version or a missing attestation is refused.
POST /api/companion/meetings/{id}/audio?channel=call|mic&seq=NRaw 16 kHz mono s16le PCMPer channel, in order; 409 {expected_seq} on a gap; a resent chunk already stored answers 200.
GET /api/companion/meetings/{id}/replies?after=NBetty’s lines for the window, the consent notice first.
POST /api/companion/meetings/{id}/endEnds the meeting. “betty leave” said in the call ends it too.
  • Auth. X-Hive-Companion-Token (per user, revocable, companion routes only). On hosted hives the hub’s auth gate waves /api/companion/ through to the spoke, which does the real check: the same pattern as the phone widget path (spokeWidgetSummaryPath in src/pkg/hub/saas_auth.go).
  • Quorum. The companion adapter segments each channel with an energy VAD, transcribes with the shared transcription service, and emits ordinary segment events (speaker is the user for mic, “Call” for call). chat.post queues the line for the companion window and reports it confirmed once the companion has fetched it, which is what satisfies rule 1. say answers platform_unavailable in v1.
  • Open in the contract. quorum.contract/v1 has no tool that carries audio from hive into Quorum; the companion platform adds one. Its name and framing are Quorum’s to fix, and hive’s routes forward to it.

7. Distribution (owner decision)

  • One Tauri codebase in a new repository, hivecommons/quorum-companion (Apache-2.0, DCO); macOS first, then Windows, then Linux.
  • No Apple Developer ID. Builds are ad-hoc signed (codesign -s -) and shipped as a GitHub Release .dmg or .zip and a Homebrew cask in hivecommons/homebrew-hive.
  • First launch needs System Settings, Privacy & Security, Open Anyway, then the Screen & System Audio Recording and Microphone permissions. Homebrew 5 removed --no-quarantine, so the cask documentation shows the Open Anyway step.
  • An ad-hoc signature has no stable designated requirement, so macOS may forget the audio-capture grant after each update. Mitigation: keep the bundle id (dev.hivecommons.quorum-companion) and the install path fixed; an optional self-signed certificate (below) gives a stable requirement.

8. Prototype PR set

Tracked in #18596:

  1. quorum: the companion platform (audio ingest, VAD, transcription, reply outbox), the contract and tests.
  2. hive: the companion meeting platform, the Quorum client calls, the /api/companion/* routes and their documentation.
  3. hive: the per-user companion token and the hub passthrough for /api/companion/.
  4. hivecommons/quorum-companion (new repository): Tauri macOS v1 (call detection, consent prompt, ScreenCaptureKit capture of the call app and the microphone, streaming, replies window), ad-hoc signed CI builds.
  5. homebrew-hive: the cask, once the first release exists.

Step 2 depends on step 1’s contract; step 3 can land beside it.

9. Owner steps

No secrets go in issues or pull requests.

  • Nothing to buy for the companion.
  • Optional, for macOS permissions that survive updates: create a self-signed code-signing certificate and add it to hivecommons/quorum-companion as the Actions secrets MACOS_SIGNING_P12 and MACOS_SIGNING_P12_PASSWORD. Until then releases are ad-hoc signed.
  • For phone dial-in later: a telephony account (Twilio or equivalent) and meetings hosted on a paid Zoom plan.

10. Sources