ElevenLabs' Hosted MCP Server Lets Claude Manage Your Voice Agents, Deletions Included

A production voice agent handling real customer calls can now be inspected, reconfigured, or deleted from inside a Claude chat window, no dashboard tab required. That is what ElevenLabs shipped this week with its hosted MCP server, a remote connector that exposes agent management as tools an AI assistant can call directly. It picked up more than 1,500 likes and 160 reposts on its launch announcement, day-one traction rather than a tracked GitHub trajectory, since this ships as a hosted connector rather than something you clone and run yourself.

What ElevenLabs' hosted MCP server exposes

ElevenLabs already had a local, open-source MCP server that bridged Claude and Cursor to its audio platform, text-to-speech, transcription, voice cloning, the kind of thing you'd run to generate a voiceover from a Claude session. The hosted MCP server is a different, narrower tool aimed at a different job: managing the ElevenAgents you've already built, the conversational voice and chat agents ElevenLabs customers deploy for things like outbound calling or support.

Once connected, Claude can create and configure a new agent, including its system prompt, voice, language, and greeting message; update any setting on an existing one; list and compare agents side by side; pull recent conversations and read full transcripts to see how an agent actually performed; estimate what a config change would do to LLM usage and cost before you commit to it; duplicate an agent; check a knowledge base's size; and generate short-lived text-to-speech links. It can also delete an agent outright.

Connecting it takes no local setup

Unlike the original ElevenLabs MCP server, which required cloning a repo, installing Python dependencies, and running a local process, the hosted version needs nothing installed on your machine. In Claude Desktop, it's Settings, Connectors, search for ElevenLabs, and complete OAuth authentication against your ElevenLabs account. From that point, Claude has read and write access to whatever agents that account can see, governed entirely by the OAuth scope you granted, which you can revoke at any time from the same settings screen.

That's the real dividing line between who each server is for. The local server suits a developer building something new with ElevenLabs' audio APIs. The hosted server suits someone who already has agents running in production and wants a faster way to check on them, tune a system prompt, or see why a conversation went sideways, without opening a separate dashboard for every question.

What using it looks like

In practice, the pitch is collapsing questions that used to require a dashboard into a single sentence. Ask Claude to pull up an agent's last ten conversations and summarize where it deviated from its intended tone, and it reads the transcripts directly. Ask what switching an agent from one model to a cheaper one would do to monthly LLM spend, and it runs the cost estimate before anything changes. Ask it to spin up a near-duplicate of an existing agent with one different greeting for A/B testing, and it duplicates and edits rather than requiring you to rebuild the config from scratch.

The honest limitation: approval isn't verification

ElevenLabs' own documentation is direct about the sharpest edge of this: deleting an agent is destructive, and it tells users to review tool calls in their MCP client before approving them. That warning is doing real work, because the failure mode here isn't obscure. A confirmation screen showing that a tool call was approved only proves the action ran, not that it did what you actually intended, and voice agent configuration has enough moving parts (system prompt, voice, knowledge base, greeting) that a single miscommunicated instruction can land somewhere you didn't mean it to.

ElevenLabs' answer to that gap is a separate agent testing framework that lets a team simulate a conversation before a change goes live, checking whether the agent responds as expected and invokes the right tools, rather than finding out from a live customer call. That testing step is not automatic. It's a second tool you have to remember to use, which means the actual safety of this setup depends on a team building the habit of testing before approving, not on the MCP server enforcing it for you.

Who the hosted MCP server fits, and who should wait

Teams already running ElevenLabs voice or chat agents in production, and who want a faster loop for reviewing performance and iterating on configuration, get real value here: fewer dashboard trips, transcripts and cost estimates available inline, all without opening a second app. Anyone still building their first agent is better served starting in the ElevenLabs dashboard directly, where the visual editor makes it easier to see the full shape of a configuration before handing edit access to a chat interface that can delete what you built with the same ease it created it.

Sources: ElevenLabs, hosted MCP server documentation; ElevenLabs blog, "Introducing the ElevenLabs MCP server"; The New Stack, "Claude can now delete your production voice agent from a chat window".