v3.0.0-alpha.1

First v3 alpha. Breaking changes; see the v3 upgrade guide in the wiki.

Install: npm i @elevenlabs/elevenlabs-js@alpha / pip install --pre elevenlabs

Renamed

• conversationalAi / conversational_ai → agents; the old agents
  sub-resource is flattened onto the group (client.agents.create).
  API URLs are unchanged (/v1/convai/*). Per-agent LLM usage stays
  nested at agents.agents.llmUsage.
• Per-endpoint request/response helper types are now verb-first
  (AgentsListRequest → ListAgentsRequest).
• Python: OnPremInitiationData → OrchestratorConfig,
  on_prem_config → orchestrator_config, on_prem_conversation_url →
  url (now validated as ws/wss). Wire format unchanged. Closes #840.
• Python: internal speech_engine.types → speech_engine.session_types.

WebSockets — now generated clients

• Hand-written realtime clients removed in favour of generated ones:
  speechToText.realtime, textToSpeech.realtime,
  textToSpeechMultiContext, textToDialogue(.MultiContext),
  translate.realtime, dubbing.realtime, speechEngine.upstream.
• Dropped: URL-based streaming (the ffmpeg-backed url option) and the
  RealtimeConnection / RealtimeEvents / AudioFormat /
  CommitStrategy root exports.

Removed (deprecated in v2)

• voices.getAll → voices.search
• textToVoice.createPreviews → textToVoice.design
• conversationalAi.addToKnowledgeBase → agents.knowledgeBase.documents.*
• agents.simulateConversation (+stream) → agent-testing endpoints
• mcpServers.approvalPolicy.update → agents.mcpServers.update
• usage.get → workspace analytics usage queries
• dubbing.transcript.get → dubbing.transcripts.get
• all 13 dubbing.resource.* methods → Dubbing Studio endpoints
• cloudStorageUrl on speechToText.convert → sourceUrl

Runtime / packaging

• JS: requires Node 18+; native fetch (node-fetch dependency gone).
• JS: dual CJS/ESM build with a full exports map; deep imports of
  dist/... must move to the documented subpaths.
• JS: request properties explicitly set to undefined are now omitted
  rather than sent as null.
• JS: ws stubbed in the browser field so bundlers don't include it.

Other

• JS: Speech Engine attach() takes one options object —
  engine.attach({ server, path, ...callbacks }).
• JS: the never-exported conversationalAi/conversation deep-import
  classes are deleted; use @elevenlabs/client for voice conversations.
• Both repos ship an elevenlabs:sdk-migration agent skill for
  automated v2 → v3 migration.