A platform for community discussion. Free, open, simple.
LiveKit agent participation
Product contract
live_livekit_agent_enabledcreates one negative-ID bot,livekit_agent_bot. Disabling preserves the account; enabling again reuses it. An existing unrelated account with that username is never adopted.- The invite modal lists the dispatch names of agents deployed to the LiveKit Cloud project (read through the Cloud agents catalogue with the project credentials, cached briefly) and falls back to a typed dispatch name when nothing is listed. The feature uses the existing LiveKit Cloud project credentials and dashboard deployment.
- Members of
live_livekit_agent_invite_allowed_groups(admins by default, admins always included) explicitly invite the bot through a public Live room's menu, including the widget menu. A human must already be present, and the call must already use LiveKit. - The bot speaks and listens. Room managers can kick it without a timed exclusion; inviters can immediately invite it again. Private rooms and mesh calls are outside this feature's scope.
- Bots leave when the last human leaves and do not contribute to human capacity, attendance, co-presence, badges, or statistics.
- No integration records, custom worker credentials, room allowlists, role fields, or exclusion management are needed. The single bot ID uses PluginStore.
Authorization and lifecycle
The invite endpoint uses normal session authentication, CSRF protection, and Guardian invite-group/public-room checks. The dispatcher independently requires the feature, a configured Cloud project and a valid invitation name, an active bot, human presence, and a LiveKit transport pin.
Every invitation writes a new opaque Redis admission proof. An authenticated provider snapshot must match the dispatch ID, room, agent name, and metadata and report exactly one running job identity. The participant must have the provider's agent kind. Neither a participant's claimed metadata nor its display name is sufficient. Clients receive the verified identity mapping in roster metadata.
Provider permission updates precede roster admission. A Redis compare-and-set checks that the authorization proof remains current when committing presence; a concurrent kick or changed setting cannot restore a stale session. Signed webhooks trigger reconciliation rather than importing event payloads as members. The scheduled sweep reconciles missed events and expires absent agents.
Kicking revokes local authorization before cancelling the provider dispatch. Cleanup records remain separate from admission proofs, allowing deletion retries and recovery of dispatch creation whose response was lost. Room cleanup resolves session keys from the room's dispatch records and the shared bot, without scanning Redis. The scheduled sweep also retries provider cleanup for deleted rooms using their retained IDs; successful cleanup removes them from the provider-room index. A new invitation supersedes earlier sessions. Disabling the feature revokes current sessions. Last-human cleanup also handles pending dispatches.
LiveKit supplies the initial permissions for managed agents. Discourse updates these before admitting playback, but this does not prevent provider access before reconciliation. Use trusted dashboard deployments. Audio conversation is supported; provider data publishing is disabled. Outages can delay remote disconnection.
Hosting and testing
LiveKit's Agents framework supports Cloud and self-hosted servers, while Agent Builder and managed hosting belong to Cloud. This implementation targets Cloud only. The operations guide documents setup, manual testing, and automated test commands. Backend tests stub provider responses; real agent audio, deployment readiness, and provider logs require manual validation.