Let Junior join Google Meet calls as a participant

Start with packages/junior-google/src/config.ts, the plugin hooks in packages/junior-plugin-api/src/hooks.ts, and the Google Meet API documentation linked in the issue. The proposal is a brainstorm spanning post-call processing, live media, and speaking, with unresolved consent, hosting, identity, and Linear assignment questions. A first step is to choose and scope one phase; done criteria depend on that design decision.

Junior can schedule Calendar events with a Meet link, but it cannot take part in the call. The idea: people invite Junior to a meeting, such as a weekly Linear backlog triage. Junior follows the conversation and sometimes looks at the shared screen. It answers when people talk to it, and it acts: it adds discussion context to the Linear issue on screen, or starts a first draft of a ticket that someone assigns to it.

What exists today

• packages/junior-google (#2063) acts as [email protected] and supports Calendar only. Scopes: calendar.events.freebusy, calendar.events.owned, calendar.events.readonly (src/config.ts). createCalendarEvent always adds a Meet link. There are no Meet or Drive scopes.
• Plugins can register webhook routes, publish Events, and add tools (packages/junior-plugin-api/src/hooks.ts). Linear already uses this for issue webhooks.
• Linear publishes only issue.created (packages/junior-linear/src/events/issue.ts). There is no assignment Event, so "assign Junior in Linear, then Junior starts work" cannot be an Event automation today.
• A Conversation Location is a Slack place only (packages/junior/src/chat/conversations/sql/location.ts). Steering messages already let new input join an active turn. This fits a stream of meeting utterances.
• @sentry/junior-tts makes WAV files with Gemini TTS. It does not stream.
• Junior runs on serverless functions and Sandboxes. Nothing in the runtime can keep a WebRTC connection open for a full call.

Google Meet APIs (checked 2026-10-12)

• Meet REST API and Workspace Events API (GA). Events for conference start/end, participant join/leave, and transcript/recording/smart-note fileGenerated. conferenceRecords.transcripts.entries gives text per speaker. This works only after the call, and only when transcription is on. Meet deletes transcript entries 30 days after the call. Files go to the organizer's Drive. artifacts, events
• Meet Media API (Developer Preview since 2025-02-24). Real-time audio and video per participant over WebRTC. Screen share is a video stream. It only receives: it cannot speak, send chat, or show UI. The scopes (meetings.conference.media.readonly or .audio.readonly/.video.readonly, plus meetings.space.read) are restricted and need verification and a security assessment. A host or same-org participant must consent, and any participant can stop it. It is blocked for encrypted or watermarked meetings and for room hardware. The client must use libwebrtc with AV1/VP9/VP8 and send metrics. overview, get started, release notes
• Meet Add-ons SDK (GA). A web app in the Meet side panel or main stage. It could show Junior's replies in the call. It does not give media.
• Google has no API to post to Meet chat or to speak in a call.

Proposal (brainstorm, as requested)

Phases, from lowest to highest risk:

1. After the call (GA APIs only). Junior organizes or is invited to the event and turns on auto-transcription for the space. On transcript.fileGenerated, Junior reads the entries, starts a Conversation in a linked Slack channel, posts a summary, and proposes or does follow-ups: Linear comments, new tickets, draft PRs. This gives much of the triage value without live media.
2. Listen live (Media API). A long-running media worker outside Vercel functions joins with the Media API. It streams audio to speech-to-text and uses the per-participant source id to name speakers. When someone says "Junior …", it sends that utterance and the recent transcript to the meeting Conversation as a steering message. Replies go to a Slack thread or an add-on side panel. For the screen, the worker takes a frame from the screen-share stream on request ("Junior, look at this") or when the screen changes, at a low rate. The turn gets the frame as an image, so Junior can find the Linear issue on screen.
3. Talk back. This needs a real participant. Options: a headless Chrome bot signed in as junior@ (TTS into a virtual mic, read and send Meet chat, take screenshots), which is fragile and has unclear ToS. Or a hosted meeting-bot vendor, such as Recall.ai, which sends meeting data to a third party.

Open questions

• Consent and recording policy for internal and external attendees (Legal/IT). Retention of transcripts and screen frames.
• Does an internal-only OAuth app skip restricted-scope verification? Does the Media API preview need every participant enrolled? Both unknown.
• Where the media worker runs (always-on vs. one per meeting), and the cost of speech-to-text and vision.
• Actor and credentials: map the Meet speaker to a Junior User/Identity. Linear MCP uses per-user OAuth, so whose account acts on a spoken instruction?
• Linear needs an assignment Event for the "assign Junior a ticket" flow.
• Keep the Slack thread as the Location, or add a meeting Location? Meet-specific logic must stay in junior-google (policies/provider-boundaries.md).
• Disclosure: anyone in the call can ask about anything Junior's account can read (same risk as #2063).

Raised by Daniel Griesser.

Related: #2063, #2057

via daniel.

--

View Junior Session [Sentry]