ElevenLabs Connects More Than 50 Media Models to Its Hosted MCP, Letting Agents Generate Complete Assets in a Single Session

ElevenLabs expanded its hosted MCP server on September 14, extending its previously ElevenAgents-focused tools to include ElevenCreative capabilities for speech, transcription, dubbing, sound effects, music, image, and video generation. The company says a single connection can access more than 50 models. In supported assistants, users can generate a script and then create narration, a soundtrack, cover art, or a video without building a separate API integration for each media type.

The connection endpoint is `https://api.elevenlabs.io/v1/mcp`. Claude Code can add it with a single MCP configuration command, while ChatGPT, Claude, and Cursor provide directories or configuration interfaces. Authentication now uses a one-time OAuth flow, eliminating the need to place a long-lived API key in the agent environment. Generated results are saved in the ElevenCreative workspace, where they can be further edited and exported in Studio. This differs from the trust model of older local MCP servers: the server, tool catalog, and model routing are all hosted by ElevenLabs.

The technical change amounts to more than adding a few generation buttons. MCP clients can discover tools at runtime, allowing agents to select media models based on the task, preserve session context, and regenerate only the narration, music, or an individual shot. For automated content pipelines, workflows that previously required multiple SDKs, file handoffs, and polling jobs can now be orchestrated through a single tool interface. However, ElevenLabs has not yet published quantitative results for end-to-end latency, failure retries, model-version pinning, or cross-model consistency.

Deployers should pay particular attention to permissions and cost controls. According to the company’s documentation, OAuth access can range from read-only monitoring to full agent management, while the same connection can now trigger computationally expensive video generation. Teams should not treat the absence of API keys as making the integration low-risk. They should still restrict available tools, require approval for high-cost or state-changing calls, and record the model used, prompts, source assets, and output versions. Feature availability may also roll out gradually across clients, so teams should verify each client’s tool catalog before production deployment.