ElevenLabs published a hosted MCP server available in Claude: sign in once with OAuth, then configure, monitor and improve voice and chat agents in natural language. The server is hosted by ElevenLabs, so there is nothing external to run or maintain, there are no API keys to manage, and the connection can be revoked from either the MCP client or ElevenLabs account settings. Claude-side settings control what the user's agents and a team's Claude sessions may access.
The interesting decision is scoping. ElevenLabs says it deliberately curates the tool set so Claude has a clear set of actions it can combine into full jobs, rather than exposing the whole API. The published examples are concrete: surface topics from recent conversations for discussion, duplicate an agent for a new market while updating voice and first message, compare configurations across agents and flag inconsistencies, and estimate an agent's expected LLM usage and cost before a change ships. The earlier 2025 local MCP server, which runs with the user's own API key, remains available for developer workflows.
A hosted server with OAuth moves the credential problem off the user's machine, and a curated tool list is a scope decision made by the vendor rather than by whoever wires up the client. That is the same administrative surface described in WorkOS Management MCP making admin state agent-addressable, and the MCP hub tracks how those surfaces are scoped.
This is a launch post describing the hosted tool surface without publishing the OAuth scopes or an audit format for actions taken. The condition to watch is whether tool calls that change a live agent's configuration are separately logged, because natural-language configuration is only reviewable if the resulting change is recorded as a diff.