Developers can now ask Claude to inspect a production voice agent, modify its system prompt, swap voices, or estimate costs for using a different language model without opening the ElevenLabs dashboard. On Monday, ElevenLabs introduced a hosted MCP connector that grants Claude read and write access to conversational agents built with ElevenAgents. The integration allows for agent creation, configuration comparison, and calculation of expected LLM usage and expenses before implementing any changes.
The connector enables Claude to update an agent's system prompt, language, voice, and opening message, retrieve transcripts, explore conversation topics, check knowledge base size, generate sample speech, and even delete agents. Developers install the connector from Claude's directory and authenticate via OAuth, eliminating the need to run a local server or paste API keys. For example, a developer can ask: "What would my checkout agent cost per conversation with Gemini 2.5 Flash instead of GPT-4o?" According to ElevenLabs, the new server focuses on managing agents that already exist in an ElevenLabs workspace, while the original open-source server released in April 2025 was designed for Claude Desktop and other MCP clients to generate speech, clone voices, or transcribe audio locally.
Because some tools can alter or remove production agents, teams are offered two layers of access control. Administrators can disable tools across an organization, and users can set stricter limits for their own sessions. ElevenLabs specifically warns that deleting an agent is destructive and recommends reviewing tool calls before approving them. The report notes that the access-control pattern mirrors what other platforms have landed on when opening production systems to AI agents, citing GoDaddy's quote-then-execute model with idempotency keys and AWS's Dogwood policy engine as examples. ElevenLabs' two-layer model sits somewhere between these approaches.
The connector introduces new workflow challenges because a successful tool call only shows that the action ran, not that it produced the intended result. A developer could ask Claude to trim a support agent's system prompt to save tokens, but if the revision accidentally removed a line telling the agent to escalate billing disputes to a person, the shorter prompt might still look fine during review. The report explains that the confirmation screen tells the team the change was approved, but it can't show whether the prompt still works as expected or a model switch has changed the agent's tool calls. ElevenLabs addresses these risks with an agent testing framework that lets teams simulate a conversation before deployment and see whether the agent responds as expected, including when it invokes a tool, with the same tests available through the CLI or API. The platform also offers opt-in agent versioning, which lets developers save configuration changes on separate branches and route a portion of production traffic to them for gradual rollouts or A/B testing, though versioning can't be disabled once enabled.
For teams that want agent configurations in a repository, the ElevenLabs CLI can pull and push agents as code, with a documented CI/CD workflow that includes a dry run before deployment and a status check afterward. The hosted MCP documentation explains how Claude can change an agent, but not whether that update enters ElevenLabs' testing and versioning workflow or goes straight to the current configuration, and it's unclear whether a team could reverse the change from Claude. For now, teams need to choose between the conversational speed of Claude for inspecting and modifying configs and the safety of CLI and API pipelines for automated testing and versioned rollouts. The report concludes by noting that "the barrier between asking an AI to analyze infrastructure and asking it to alter it is disappearing." Organizations will need to determine whether conversational convenience or systematic validation better serves their operational discipline as model capabilities continue to outpace established developer workflows.