ElevenLabs MCP Server

Official

Give your assistant a voice — text-to-speech, voice cloning and audio tools from the ElevenLabs API.

Local serverstdioPythonMIT 1.2k

What is the ElevenLabs MCP server?

The ElevenLabs MCP server turns text conversations into audio production sessions. Backed by the same API that powers most of the convincing AI voices you hear online, it lets your assistant generate speech in hundreds of voices and dozens of languages, design entirely new voices from text descriptions, clone voices from samples, transcribe audio, and even isolate speech from noisy recordings.

The conversational framing genuinely fits this domain. Voiceover work is iterative — "warmer, slower, more conviction on the second line" — and iterating through an assistant that regenerates the clip in seconds beats round-tripping through a studio interface. Scripts written and voiced in the same conversation collapse a whole workflow: draft the ad copy, hear it, revise both words and delivery, export.

Practical touches worth knowing: output files land in your Desktop/ElevenLabs folder by default, voice-design tools let you conjure "a gravelly narrator with a slight Irish accent" without auditioning a library, and transcription closes the loop for podcast and interview workflows. Costs follow your ElevenLabs plan's character quota — generation is metered, so bulk audiobook ambitions need a paid tier, while the free tier covers experimentation.

Setup is one API key via uvx. If your work touches spoken audio at all — product demos, localised voiceovers, podcast tooling, accessibility — this is the most capable audio server in the MCP ecosystem, from the company that set the current bar for synthetic speech.

Getting good takes faster

  • Pin a voice ID as soon as you like a read. Passing the same voice_id through the rest of a script keeps the character consistent instead of drifting line to line.
  • Punctuation is stage direction. An ellipsis buys a pause, a comma shortens one, an em dash lands as a beat — rewriting the copy usually beats re-rolling the same sentence hoping for a different delivery.
  • Multilingual v2 handles accented and non-English text with the most nuance; the Flash models trade some of that for latency, which is the right trade when you are auditioning twenty variants of one line.
  • Chunk long scripts by paragraph and stitch the results. Single enormous requests are expensive to redo and prosody tends to wander by the end of them.

What you can do with it

Script-to-voiceover in one session

Write, voice, critique and re-voice marketing or video copy iteratively.

Localised audio

The same script rendered across languages and voices for global content.

Podcast tooling

Transcribe episodes, clean noisy recordings, generate intros.

Available tools

ToolWhat it does
text_to_speechGenerate speech from text in a chosen voice
search_voices / get_voiceBrowse and inspect the voice library
voice_cloneClone a voice from audio samples
text_to_voiceDesign a new voice from a text description
speech_to_textTranscribe audio files
isolate_audioSeparate speech from background noise
play_audioPlay generated audio locally

How to install the ElevenLabs MCP server

{
  "mcpServers": {
    "elevenlabs": {
      "command": "uvx",
      "args": ["elevenlabs-mcp"],
      "env": { "ELEVENLABS_API_KEY": "your-api-key" }
    }
  }
}

Configuration

An ElevenLabs account and API key; character quotas follow your plan. Python with uv.

VariableDescriptionRequired
ELEVENLABS_API_KEYAPI key from the ElevenLabs dashboardYes

Example prompts to try

  • Voice this 30-second ad script with a warm female narrator, then try a British male read.
  • Design a voice: elderly storyteller, slow cadence, slight rasp — and read this fable with it.
  • Transcribe this interview recording and pull the five best quotes.

Frequently asked questions

The server is free and open source; generation spends your ElevenLabs character quota. The free tier suits experiments — production volumes need a paid plan.