Google Meet tool and modes

Tool actions for agents, session status fields, and the agent and bidi talk-back modes. Part of the Google Meet plugin guide.

Tool

Agents use the google_meet tool:

{
  "action": "join",
  "url": "https://meet.google.com/abc-defg-hij",
  "transport": "chrome-node",
  "mode": "agent"
}

test_speech always forces mode: "agent" or "bidi" and fails if asked to run in mode: "transcribe", because observe-only sessions cannot emit speech. speechOutputVerified requires both fresh realtime output bytes and a matching non-silent waveform on the native virtual microphone capture path during that output. With generated input commands, participant input is captured separately from browser playback; explicit input commands retain their configured capture path. A reused session's older output or loopback signal does not count, and sink-byte growth alone does not report verified speech. This verifies local microphone injection; a second participant is needed to confirm remote audibility.

For Chrome transports, leave keeps a reused user-owned tab open after clicking Meet's Leave call button. Tabs opened by OpenClaw are closed after departure.

Use transport: "chrome" when Chrome runs on the Gateway host, transport: "chrome-node" when it runs on a paired node. In both cases the model providers and openclaw_agent_consult run on the Gateway host, so model credentials stay there. Agent-mode logs include the resolved transcription provider/model at bridge startup and the TTS provider/model/voice/output format/sample rate after each synthesized reply. Raw mode: "realtime" is still accepted as a legacy compatibility alias for mode: "agent", but it is no longer advertised in the tool's mode enum.

create with an API-backed room and explicit access policy:

{
  "action": "create",
  "transport": "chrome-node",
  "mode": "agent",
  "accessType": "OPEN"
}

Ending a known room's active conference:

{
  "action": "end_active_conference",
  "meeting": "https://meet.google.com/abc-defg-hij"
}

Listen-first validation before claiming a meeting is useful:

{
  "action": "test_listen",
  "url": "https://meet.google.com/abc-defg-hij",
  "transport": "chrome-node",
  "timeoutMs": 30000
}

Speaking on demand:

{
  "action": "speak",
  "sessionId": "meet_...",
  "message": "Say exactly: I'm here and listening."
}

status includes Chrome health when available:

Native participation requests

Call participation_context with sessionId before using participate. It lists only capabilities available for the current tracked browser session; an empty list means no native action is available. A platform must provide the native implementation before the runtime advertises its capability. Twilio does not support browser participation actions.

Pass the advertised action object in participationAction and reuse the same requestId when checking an unclear response. Reusing that ID never repeats the effect. An uncertain result means the effect may have occurred; do not generate a new ID to retry it. Stored results remain readable after leaving, without allowing any new action on the closed session.

Cancellation is best effort for browser actions already dispatched. If the meeting ends or source authority changes during delivery, the effect may occur before the runtime detects that change. An uncertain result is not confirmation that the action was cancelled; do not retry it with a new request ID.

Incoming chat or caption requests must keep the sourceId issued by their observer. A tool caller must not remove it to turn meeting input into an unrelated operator command. Sources expire after two minutes from their first observation; interim text, edits, own echoes, and page changes invalidate old references. Edits retain their original order and do not become fresh invitations. Direct operator commands may omit sourceId.

Participation context retains at most 1,024 live sources, using their original observation order for capacity decisions. At capacity, rereading older history does not displace newer retained sources. A newer eligible observation replaces only the oldest retained source. Unchanged retained sources keep their references and live guards; rereading them does not extend the two-minute lifetime.

Retained caption rows also carry independent observation provenance, including the observed speaker label and native self/other/unknown marker. Equal text does not transfer those facts between participants. Historical, interim, own-echo, and capacity-ineligible rows retain their provenance even when they have no actionable source. A duplicate DOM row disappearing does not finalize another live copy. The latest retained snapshots are not a complete journal of intermediate edits.

A rejected request may return correctionOf. It permits one corrected request with a new requestId, that exact correctionOf, and the same source and action type. It does not permit retrying an uncertain effect or changing who authorized it. Attempts and results use the existing SQLite plugin state store. History is bounded: at capacity, the oldest closed-session rows are reclaimed. Active claims never expire or get evicted to admit another action; discarded closed requests remain inactive and cannot run again. A known pre-insertion capacity rejection permits one fresh atomic admission attempt after bounded cleanup, including when another request reclaimed the same rows. Unknown write outcomes are never retried automatically.

Agent and bidi modes

agent mode: the realtime transcription provider hears meeting audio, final participant transcripts route through the configured OpenClaw agent, and the answer is spoken through regular OpenClaw TTS. Nearby final-transcript fragments are coalesced before the consult so one spoken turn does not produce several stale partial answers. Isolated browser input remains available during TTS; only legacy transports with mixed loopback input retain playback and transcript echo suppression.

bidi mode: the realtime voice model answers directly and delegates deeper reasoning, current information, or normal OpenClaw tools to the configured agent. Providers with function tools use openclaw_agent_consult; GPT-Live uses native delegation through the same Gateway voice runtime as Discord and Talk. Both paths preserve meeting context and realtime.toolPolicy. The resulting answer is spoken by the realtime voice model.

For GPT-Live with Cove, use the Live configuration. Live owns response timing and interruption: incoming participant audio stays open during playback, and OpenClaw does not add a local barge-in detector or wait for a response-completion event to play short replies. Stopping the meeting or replacing the provider session cancels active delegation and ignores late results.

By default consults run against the main agent; set realtime.agentId to point a Meet lane at a dedicated agent workspace, model defaults, tool policy, memory, and session history. Agent-mode consults use a per-meeting agent:<id>:subagent:google-meet:<session> session key so follow-up questions keep meeting context while inheriting normal agent policy. When an agent calls google_meet in agent mode, the consultant session forks the caller's current transcript before answering participant speech; the Meet session stays separate so meeting follow-ups do not mutate the caller transcript directly.

realtime.toolPolicy controls the consult run:

The consult session key is scoped per Meet session, so follow-up consult calls reuse prior consult context during the same meeting.

Force a spoken readiness check after Chrome has fully joined:

openclaw googlemeet speak meet_... "Say exactly: I'm here and listening."

Full join-and-speak smoke:

openclaw googlemeet test-speech https://meet.google.com/abc-defg-hij \
  --transport chrome-node \
  --message "Say exactly: I'm here and listening."