ElevenLabs MCP: one OAuth, 50 models, scope the tools

ElevenLabs shipped image and video generation into its MCP connector on September 14. That completes a set: speech, transcripts, dubs, music, sound effects, images and video, all reachable from whatever assistant your team already types into. One OAuth sign-in, 50+ models, no API keys. The install takes about thirty seconds, which is exactly the problem worth thinking about before Friday.

What actually happened

Per ElevenLabs' announcement, you find ElevenLabs in the connector directory of Claude, ChatGPT or Cursor, click connect, and sign in with your ElevenLabs account. No server to run, no keys to rotate by hand. It is live in Claude, ChatGPT, Cursor, Grokbot and Hermes.

The tool surface: Text to Speech, Scribe for speech to text, dubbing, music and sound effects, image generation, and video with lipsync. They chain — generate an image, edit it, animate it into video, add lipsync, without leaving the conversation. Every generation lands in your ElevenCreative workspace, where you open it in Studio to finish.

Two admin controls matter. Workspace admins decide which tools the connection can call. And data residency is selected at connection time — but ElevenLabs lists residency as an enterprise-only feature.

Why an MCP media connector matters for your business

The credit meter moved into the chat window. Text-to-speech is cheap. Video generation is not. Once the connector is live, anyone who can type a sentence can start a render, and the spend shows up in a creative workspace rather than in the engineering budget where someone watches it. Set the tool allowlist first and widen it on request — not the other way around.

An OAuth connector is a standing grant, not an action. It keeps working after the person who installed it changes roles or leaves. Treat the connector list on your Claude, ChatGPT and Cursor workspaces the way you treat API keys: inventoried, owned by a name, reviewed on a schedule. We have made the same argument about putting a policy in front of MCP writes — the principle does not change because the tool makes audio.

Residency behind an enterprise tier is a real constraint. If you run support calls, medical intake, or anything covered by a client contract, "where is this audio processed" is not a nice-to-have. On a lower tier you do not pick. Check your plan against your customer commitments before you pipe a single recording through Scribe.

Assets landing in a vendor workspace need an export path. ElevenCreative is convenient and it is theirs. If a campaign's voiceovers, music beds and lipsynced video only exist there, you have a migration project waiting. Pull finished assets into storage you control as part of the workflow, not as a cleanup task later.

Sources: ElevenLabs, ElevenLabs MCP docs.