Publication date: 15 September 2026
ElevenLabs has expanded its hosted MCP so assistants can directly generate speech, transcripts, dubbing, music, sound effects, images and video. The company published the change on 14 September and says it is live now in ChatGPT, Claude, Cursor and other compatible clients.
This is a material deployment change rather than a new base model. For marketing and creative operations teams, it removes much of the API glue between a conversational brief and generated assets. It also moves generation permissions into an agent-facing connector.
The new route turns an assistant thread into a creative production surface
ElevenLabs says one hosted connector now exposes more than 50 models. Users sign in once with OAuth and do not need to run an MCP server or manage an API key. Generated assets land in the user’s ElevenCreative workspace, where they can be refined in Studio.
The workflow can chain different modalities. A brief can become a script, voiceover, music bed, sound effects, images and video without switching applications. Scribe transcription is available through the same connector, including speaker labels and timestamps across 99 languages according to the launch post.
| Route | Best fit | Operational trade-off |
|---|---|---|
| Hosted ElevenLabs MCP | Human-led work inside ChatGPT, Claude, Cursor and similar assistants | Fast setup, OAuth and shared workspace permissions |
| Direct ElevenLabs API | Production applications and deterministic pipelines | More engineering work, but explicit endpoint and program control |
| ElevenCreative UI / Flows | Visual creative assembly and manual refinement | Strong editing control, more context switching |
The MCP is strongest when a person already works in an assistant and wants the agent to orchestrate multiple ElevenLabs capabilities. It does not replace the API for back-end production systems.
Pricing stays tied to the ElevenLabs credit pool
ElevenLabs does not announce a separate MCP connector fee. Its public pricing uses one shared monthly credit pool across products. The Free plan includes 10,000 credits, Starter costs $6 per month with 30,000 credits, and larger plans increase the pool and workspace features.
Credit consumption differs by medium. The pricing page lists about 330 credits per minute for speech-to-text, 900 per minute for music, 200 per sound-effect generation and 2,000 credits per minute for automatic dubbing with a watermark. Image and video availability varies by plan, with the Starter tier adding Image & Video access.
This makes the first buyer question less about connector cost and more about generation volume. A campaign agent can trigger several expensive media operations from one conversational request, so teams should measure credit consumption per finished asset rather than per prompt.
Governance is the main reason not to enable everything by default
Workspace administrators can decide which tools the connection may call, and the launch announcement says data residency is selected when connecting. OAuth scopes access to the workspace and can be revoked without rotating API keys.
However, ElevenLabs’ broader MCP documentation shows why permission design matters. MCP capabilities can act on external systems, and Zero Retention Mode or HIPAA-compliant workspaces cannot use MCP servers in the agent integration documented by ElevenLabs. The company’s data-residency documentation also notes that storage location and processing location are not always identical unless stricter configurations are used.
The new creative MCP therefore needs the same controls as any agent that can spend money or publish assets. Read-only monitoring and generation tools should not automatically share the same approval policy.
What marketing and creative teams should test first
Start with one bounded production task, such as turning an approved 30-second campaign script into a voiceover, music bed and first-cut visual asset. Set a credit budget and require human approval before any expensive regeneration or external publication.
Compare elapsed time, number of application switches, revision count and credits consumed against the current manual workflow. For teams already using ChatGPT, Claude or Cursor, the reduction in integration work is large enough to justify a pilot. The main reason not to roll it out broadly is governance, not model quality.
What this means
Marketing, content and creative operations teams that already work inside supported assistants should test the ElevenLabs MCP now. The first workload should be a repeatable asset package with clear approval points and measurable credit consumption.
For production software teams, keep the direct API as the default deployment route. The MCP is best viewed as a human-facing orchestration layer, with strong convenience but a larger permission surface.
For more information, visit the official announcement of ElevenLabs MCP on the ElevenLabs website.
