Claude can now delete your production voice agent from a chat window

ElevenLabs has launched a hosted MCP connector, enabling Claude to interact with voice agents built using ElevenAgents. This allows developers to perform various tasks, such as inspecting agents, modifying system prompts, changing voices, estimating costs for different language models, and even deleting agents, all without needing to access the ElevenLabs dashboard.

For example, a developer could inquire about the cost of using Gemini 2.5 Flash instead of GPT-4o for a checkout agent in a chat. This capability becomes increasingly vital as cheaper models may not necessarily result in lower costs when agentic orchestration is involved. To use the connector, developers install it from Claude's directory and sign in using OAuth, eliminating the need for local server setup or sharing API keys.

The new server focuses on managing existing agents within an ElevenLabs workspace. Besides the aforementioned tasks, Claude can retrieve transcripts, explore conversation topics, check the size of a knowledge base, generate sample speech, and delete an agent. However, because some of these actions can modify or remove a production agent, access control is implemented with two layers.

Administrators can disable tools organization-wide, and users can set stricter limits for their individual sessions. ElevenLabs cautions that deleting an agent is irreversible and advises reviewing tool calls before approving them. The access control system resembles the quote-then-execute model seen in other platforms, which safeguards against irreversible actions initiated by automated agents.