At 11:07 PM (KST) on August 17, 2026, the official ElevenLabs account made an announcement with a 45-second intro video: "Introducing the ElevenLabs MCP, now available in Claude." It says you can manage the voice and chat agents your team built directly inside Claude.
According to the announcement, there are three things you can do from Claude: check recent performance, create new agents, and change settings. It also includes a feature that calculates the expected cost in advance when you switch models.
No API Key: Just Log In to Connect
In a reply posted nine hours later, ElevenLabs explained, "You just log in securely with OAuth, and there's no need to spin up a server or manage API keys," and left a link to the connection page. It added that what Claude sessions and agents can access is limited to the permissions you approve at login.
It's worth noting that this isn't the first announcement of its kind. ElevenLabs already released an open-source MCP server in April 2025. Back then, developers had to run the server locally themselves and issue and manage their own API keys. What's new is a managed connector where ElevenLabs hosts that server for you, cutting setup down to a single login.
Change Settings in Plain Language, Without a Dashboard
One reply argued that "configuring a voice agent through a natural-language interface instead of a dashboard is itself evidence that dashboards are being replaced by natural-language interfaces." Users used to the developer routine of issuing API keys reacted with "it only took until now to get here."
It Can Also Delete Your Live Agents
This connector is a write-enabled integration with permission to change and create settings, not just read them. The tech outlet The New Stack pointed out that "you can now even delete a live voice agent from a chat window." Before connecting it to a production environment, you should first review your team account's permission scope and approval process.
Questions in the replies asked whether it works with other coding agents such as Codex and exactly how it differs from the API approach, but as of August 18 (KST), no separate answer from ElevenLabs had been confirmed.