Your next voiceover could start in the same conversation where you wrote the script. A voice agent could also take shape inside the assistant you already use for work.
ElevenLabs’ hosted MCP makes those workflows possible. Its Agents quickstart supports creating and managing agents through natural-language requests in Claude or another compatible MCP client.
The more interesting change is how users access ElevenLabs during a workflow. Instead of opening another app to handle audio, users can let an AI assistant call ElevenLabs when a task needs speech, transcription, or an agent.
For ElevenLabs, that could mean reaching people who rarely open its dashboard.
ElevenLabs moved MCP from a developer project to a hosted service
MCP, or Model Context Protocol, gives AI assistants a standard way to connect with external tools. Developers can expose capabilities through an MCP server instead of creating a separate integration for every assistant.
ElevenLabs first introduced its MCP server in April 2025. The original version ran locally and required users to supply an API key. An August 17, 2026 update now points people toward its hosted replacement.
The original GitHub repository has since been deprecated. ElevenLabs hosts the server itself, while OAuth replaces the process of copying an API key into a client.
The hosted setup removes one of the more technical parts of MCP adoption. Someone who wants to change a voice agent through Claude can connect the service instead of maintaining a local server.
It also suggests ElevenLabs plans to treat MCP as part of the product rather than a developer experiment.
Your assistant can now manage more of the voice workflow
The ElevenLabs MCP page says assistants can create voice or chat agents and change their prompts or voices. Users can also compare configurations, duplicate agents, delete them, and view estimated LLM usage and costs.
The page advertises access to over 50 models across voice, music, images, and video. Earlier MCP documentation also describes speech generation and transcription.
ElevenLabs offers dubbing and sound effects elsewhere in its audio platform, although the hosted MCP page does not specify whether every individual capability is exposed through the service.
What changes is how these tools can become part of one continuous workflow. An assistant could help draft a script and then send it to ElevenLabs for narration. A team could create a voice agent, test a different prompt, and adjust its configuration from one conversation.
Those are potential workflows rather than proof that every combination will work automatically. Still, ElevenLabs now has more chances to appear inside projects where audio is only one part of the job.
MCP access does not turn a coding assistant into a voice assistant
Connecting ElevenLabs to a coding tool gives that assistant access to ElevenLabs capabilities. It does not automatically make the coding environment voice-controlled.
Generating speech and accepting spoken coding instructions are separate functions.
ElevenLabs has a different Cursor integration for that type of workflow. An ElevenLabs voice agent can launch Cursor background coding agents, check their progress, retrieve conversations, and send follow-up instructions. Coding tasks can also result in pull requests.
The Cursor integration requires a Cursor API key and a paid Cursor plan, separate from the OAuth connection used by hosted MCP.
Developers can also use the ElevenLabs CLI, which stores agents as code in version control and supports automated deployment workflows. MCP favors conversational control, while the CLI enables development teams to manage agent changes through their existing software process.
Compatibility is expanding faster than the documentation
ElevenLabs’ setup instructions name Claude and Claude Code, along with ChatGPT and Cursor. Hermes gets its own setup command. The FAQ on the same page, however, mentions only Claude and Claude Code.
The documentation leaves some uncertainty about which clients ElevenLabs officially supports.
OpenAI’s Responses API documentation confirms support for remote MCP servers, including services that require OAuth access tokens. Developers could therefore connect ElevenLabs through OpenAI’s API as well, although the integration would still require compatible authentication and testing.
The Neuron’s DevDay tutorial explored a similar model: assistants can call external software capabilities when a task requires them.
The dashboard could become secondary
ElevenLabs says authorization is scoped to a workspace. Permissions can range from read-only access to agent management, and users can revoke access later.
Those controls become more consequential once an assistant can change or delete agents. A connection used to inspect settings requires less authority than one used to manage production agents.
There is also a business tradeoff for ElevenLabs. MCP can put its tools in front of more users, while reducing the number of reasons those users need to visit ElevenLabs directly.
If more AI products adopt MCP, users may have fewer reasons to open each service separately. An assistant could call ElevenLabs for voice generation as part of a larger coding or research workflow.
For ElevenLabs, success may eventually mean its technology gets used constantly while its own interface fades further into the background.
